Advanced integrated SERS-based strategies for the early diagnosis of upper gastrointestinal cancers

Abstract

Upper gastrointestinal (UGI) cancers are associated with high incidence and poor prognosis due to their insidious onset, posing a serious threat to human health. Traditional diagnostic approaches, including endoscopic biopsy, imaging modalities, and molecular assays, frequently suffer from tedious procedures and insufficient sensitivity. Therefore, it is imperative to develop novel technologies for the early diagnosis of UGI cancers to reduce diagnostic delay, ensure diagnostic accuracy, and ultimately improve patient survival rates. Owing to its exceptional sensitivity and rich molecular fingerprinting information, surface-enhanced Raman spectroscopy (SERS) has emerged as an invaluable analytical tool for uncovering metabolic molecular alterations, thereby facilitating rapid, sensitive, and early-stage diagnosis of UGI cancers. Herein, this review not only outlines the principles and detection strategies of SERS technology, but also highlights its recent advances in the early diagnosis of UGI cancers, particularly in conjunction with other innovative technologies. Ultimately, the current challenges and future development directions are discussed, providing a valuable perspective to guide future efforts toward early UGI cancer screening.

Graphical abstract

Introduction

Upper gastrointestinal (UGI) cancers, predominantly esophageal and gastric cancers, are among the most prevalent and lethal malignancies worldwide [1, 2]. According to International Agency for Research on Cancer (IARC) data, UGI cancers were associated with over 1.47 million new cases and more than 1.1 million deaths globally in 2022, imposing a substantial burden on healthcare systems [3]. Notably, approximately 60% of UGI cancer cases are diagnosed at an advanced stage due to the lack of sensitive diagnostic methods, resulting in poor prognosis and a low 5-year survival rate [4].

Currently, the routine diagnostic approaches for UGI cancers mainly include endoscopic biopsy, imaging modalities, molecular assays, etc., as detailed in Sect. 2 [5, 6]. Endoscopic biopsy is generally the gold standard for UGI cancer diagnosis but has inherent drawbacks, such as invasiveness, high cost, and specialized operator requirements. In addition, noninvasive imaging modalities, such as magnetic resonance imaging (MRI) and computed tomography (CT), provide imaging information about only tumor invasion and lesion extent, and are inadequate for the early diagnosis of UGI cancers [7]. Furthermore, preclinical in vitro molecular assays, such as polymerase chain reaction (PCR) and enzyme-linked immunosorbent assay (ELISA), are inevitably constrained by prolonged detection times, tedious procedures, and expensive instrumentation [8, 9]. Therefore, the development of sensitive, rapid and nondestructive diagnostic approaches is critical for timely medical intervention and improved survival rates for patients with UGI cancers.

Surface-enhanced Raman spectroscopy (SERS) is an ultrasensitive, rapid, and nondestructive analytical technique that can drastically enhance the Raman signals of molecules adsorbed on active substrates [10, 11]. Notably, SERS has emerged as an attractive biophotonic tool for disease diagnosis, due to its abundant molecular fingerprinting information, simultaneous multiplex detection capability, and operational simplicity [12–14]. Recently, various SERS-based strategies have been explored and applied in UGI cancer diagnosis, rendering it more rapid and accuracy. While several reviews have mentioned the potential of SERS in UGI cancer diagnosis, they merely provide brief overviews of SERS-based diagnostic strategies for esophageal and gastric cancers separately, without a consolidated retrospective dedicated to SERS-based technologies and their integration with other novel strategies for UGI cancer screening (Table 1) [15, 16].

Table 1.

Systematic comparison of this review with published review articles in this field

Prior published reviews	Cancer types	Timeframe	Clinical diagnostic methods	Perspectives	Shortcomings
[15]	Gastric cancer (GC)	SERS-related studies spanning mainly from 2016 to 2023	Brief introduction	Discussion of various optical technologies for GC diagnosis	1) Lack of a systematic and comprehensive summary on SERS-based GC diagnostic strategies 2) Limited discussion of future developments of SERS-based diagnostic methods 3) Inadequate discussion of clinical application potential 4) Lack of in-depth investigation on integrated diagnostic platforms combining SERS and other advanced technologies
[16]	Esophagus cancer (EC)	2010-2021	Not included	Elaboration on SERS-assisted EC diagnosis (including urine, blood and tissue analyses)	1) Focus on label-free SERS analysis, with limited elaboration on labeled SERS strategies 2) Insufficient overview of AI algorithms and algorithm comparisons 3) Lack of in-depth investigation on integrated diagnostic platforms combining SERS and other advanced technologies 4) Limited elaboration on future developmental trends and clinical translation
This review	Cancer types	Timeframe	Clinical diagnostic methods	Perspectives	Contributions
	UGI cancers (GC and EC)	Spanning mainly from 2019 to 2026	Detailed summary	Comprehensive elucidation of recent advances in integrated SERS-based strategies for the early diagnosis of UGI cancers	1) Evaluations of the strengths and weaknesses of routine diagnostic approaches 2) Overview of SERS principles and a detailed introduction to SERS detection strategies 3) Detailed summary of AI-assisted SERS platforms and their performance in UGI cancer diagnosis 4) Comprehensive elaboration of innovative labeled joint diagnostic approaches, with an emphasis on the integration strategies of SERS with other cutting-edge technologies 5) Systematic assessment of challenges and future developmental trends of SERS-based UGI cancer diagnostic methods

This review comprehensively elucidates recent advances in integrated SERS-based strategies for the early diagnosis of UGI cancers. Initially, we briefly summarize the conventional diagnostic approaches and discuss their respective advantages and limitations. Subsequently, the fundamental principles of SERS and representative detection strategies are systematically outlined. Following this, we present a detailed description of artificial intelligence (AI)-assisted label-free methodologies and innovative labeled joint diagnostic approaches (Fig. 1). Conclusively, the current challenges and future perspectives are highlighted to encourage in-depth exploration of efficient SERS-based diagnostic strategies, thereby facilitating the early screening of UGI cancers and contributing to public health advancement.

Fig. 1.

Panoramic schematic illustration of novel integrated SERS-based strategies for UGI cancer diagnosis. Created with BioGDP.com

UGI cancer diagnostic methods and their limitations

Clinical diagnostic approaches

Currently, the primary clinical diagnostic methods for UGI cancers include endoscopic biopsy, routine endoscopy, and imaging modalities, all of which have inherent limitations in terms of achieving sensitive and rapid diagnosis of early-stage UGI cancers [17, 18]. For instance, conventional imaging modalities (e.g. CT and MRI) provide substantial anatomical information, such as tumor infiltration depth, invasion extent, lymph node involvement and distant metastasis. However, these techniques are primarily employed for pre-treatment clinical staging and are usually limited in achieving early-stage diagnosis due to their insufficient sensitivity [19, 20].

White-light endoscopy, which enables direct visualization of suspicious lesions, is a cornerstone technique for UGI cancer screening [21]. With the continuous technological advancements, numerous attractive endoscopic modalities (e.g., chromoendoscopy, electronic staining endoscopy, magnifying endoscopy, and confocal laser endoscopy) have been developed to detect subtle abnormalities and early-stage lesions associated with UGI cancers [22]. Moreover, plentiful high-performance AI models have been proposed to further enhance diagnostic accuracy and sensitivity [23, 24]. Nevertheless, these endoscopic examinations suffer from several unavoidable drawbacks, including invasiveness, time-consuming procedures, inadequate sensitivity, as well as frequent risks of hemorrhage and infection, which impede their implementation in routine screening [25–27].

Generally, endoscopic biopsy is widely recognized as the gold standard for UGI cancer diagnosis, as microscopic histopathological examinations enable the definitive identification of the histological subtype and pathological severity of lesions [17]. However, this approach is prone to misdiagnosis and false-negative results due to sampling errors or pathologists’ subjective interpretation [25]. Additionally, the inherent invasive nature of biopsy is potentially associated with risks of hemorrhage, perforation and even cancer metastasis [28].

Emerging diagnostic strategies

Given the imperfections of existing clinical diagnostic methods, emerging diagnostic strategies have been developed to improve the sensitivity and specificity of early-stage UGI cancer detection, including molecular assays, analytical chemistry strategies and biophotonic technologies. A comprehensive overview of these diagnostic strategies for UGI cancers is presented in Fig. 2.

Fig. 2.

A comprehensive overview of clinical and emerging diagnostic strategies for UGI cancers. Created with BioGDP.com

Molecular assays (e.g., PCR, high-throughput sequencing, and ELISA), which enable direct quantification of DNA, RNA, and protein biomarkers, are used to identify relevant molecular alterations in early-stage UGI cancers [29–31]. Although these methods have substantial advantages in the sensitive detection of characteristic biomarkers and the delineation of cancer-specific genetic landscapes, they still have limitations, including expensive instruments, meticulous procedures and rigorous experimental conditions.

Furthermore, diverse analytical chemistry techniques have been developed for the sensitive detection of biomarkers associated with UGI cancers [32, 33]. For instance, by using proton-transfer-reaction time-of-flight mass spectrometry (PTR-ToF-MS) coupled with gas chromatography-mass spectrometry (GC-MS) or liquid chromatography-mass spectrometry (LC-MS), Adam’s [34] and Han’s [35] groups demonstrated that specific volatile fatty acids and the tryptamine-to-tryptophan ratio could serve as robust metabolic biomarkers for the early screening of UGI cancers. These analytical techniques enabled the sensitive identification of early-stage metabolic alterations, yet still suffered from blemish such as high instrument costs, complicated sample preparation procedures, inadequate biological interpretation and so on.

Biophotonic technologies, such as fluorescence spectroscopy, polarized light scattering spectroscopy (PLSS), and confocal laser endomicroscopy (CLE), among others, also hold great promise for UGI cancer diagnosis [16]. For instance, a near-infrared ratiometric fluorescent probe and a novel fiber-based snapshot PLSS endoscopic system have been validated to effectively differentiate normal and cancerous tissues [36, 37]. Pang et al. [38] demonstrated that a high-resolution CLE system achieved a more comprehensive assessment of gastric precancerous lesions. Nevertheless, these biophotonic tools have distinct drawbacks. For example, fluorescence spectroscopy is constrained by unavoidable photobleaching effects and low detection throughput, whereas PLSS and CLE systems are hampered by invasive risks, complicated manipulation protocols and high-standard equipment.

As a promising diagnostic modality, liquid biopsy achieves early diagnosis by detecting tumor-derived biomarkers circulating in diverse biofluids, including circulating tumor cells (CTCs), circulating tumor DNA (ctDNA), and extracellular vesicles (EVs), among others [39]. Current CTC analysis predominantly employs PCR, single-cell sequencing, and cellular protein detection approaches [40]. In parallel, ctDNA detection primarily relies on PCR and next-generation sequencing (NGS)-based platforms, whereas EV profiling is commonly achieved via quantitative reverse transcription PCR, RNA sequencing, and ELISA [41]. Notably, the majority of these approaches are plagued by limited multiplexing capability, cumbersome sample pretreatment protocols, and prohibitive expenses for large-scale population screening [42, 43]. With the rapid advancements in liquid biopsy techniques, SERS stands out as a robust candidate for UGI cancer screening by virtue of its nondestructive detection, ultrahigh sensitivity, operational convenience, cost-effectiveness, real-time responsiveness, and multiplexing potential [44].

SERS principles and detection strategies

Overview of SERS

Raman scattering represents an inelastic light-scattering process involving a distinct change in photon energy (Fig. 3a) [45]. In 1974, Fleischmann and coworkers [46] first reported a pronounced enhancement of Raman signals from pyridine adsorbed on a roughened silver electrode surface, with this effect initially attributed to the enlarged effective surface area. Afterward, Van Duyne and Creighton elucidated that this remarkable signal enhancement arose from an elevated Raman scattering cross-section, triggered by the interaction between analyte molecules and the underlying metal surface [47, 48]. This phenomenon, defined as SERS, has evolved into a promising spectroscopic analytical technique renowned for its ultrahigh sensitivity down to the single-molecule level, distinctive molecular fingerprinting capability, rapid analytical speed and minimal sample consumption.

Fig. 3.

SERS fundamentals and substrate types. a) The energy level diagram of Rayleigh scattering and Raman scattering; b) the LSPR effects in plasmonic nanoparticles; and c) the CM effects in analyte-substrate systems. Reproduced with permission [45]. Copyright 2025, Royal Society of Chemistry. Schematic illustration of the two main categories of SERS substrates: d) noble-metal substrates and e) non-metallic substrates. Reproduced with permission [57]. Copyright 2022, Wiley. Reproduced with permission [60]. Copyright 2024, Royal Society of Chemistry

SERS enhancement mechanisms

Currently, the performance of SERS systems is governed primarily by two well-established theories: the electromagnetic enhancement mechanism (EM) and the chemical enhancement mechanism (CM) [49]. For the EM effect, a dramatic signal enhancement with an enhancement factor (EF) of up to 10⁸ can be observed, which is attributed mainly to localized surface plasmon resonance (LSPR) in plasmonic nanostructures (Fig. 3b) [50]. The nanogaps between noble-metal nanoparticles or sharp edges (termed “hot spots”) can yield intense electromagnetic fields, drastically amplifying the Raman response of entrapped analytes. The EM effect, regarded as the dominant and universal enhancement mechanism, is independent of the physicochemical properties and orientation of the analytes [51]. However, its efficacy highly relies on the distance between the analytes and nanoparticles, and the enhancement effect decays rapidly as this distance increases [52].

In contrast, the CM effect originates from specific interactions between adsorbed analytes and SERS substrates, and can be further categorized into molecular resonance, charge transfer (CT) resonance and non-resonant chemical effects (Fig. 3c) [53]. Molecular resonance and CT resonance constitute two major contributions to the CM effect. Specifically, molecular resonance occurs when the incident laser energy matches the analyte’s electronic transition energy, whereas CT resonance happens when the excitation light induces CT processes between the substrate and adsorbed analytes, on the premise of well-aligned energy levels in the analyte–substrate system. Moreover, non-resonant chemical effects, which are influenced by the adsorption environment, typically contribute only slightly to the overall CM effect [53, 54]. Characterized by unique surface selection rules, the CM effect results in the selective amplification of Raman modes corresponding to the perpendicular components of the polarization tensor for adsorbed analytes [12]. Additionally, the CM effect manifests as Raman frequency shifts, the appearance of new peaks, and intensity redistribution, arising from the interactions between the analytes and substrates [55].

SERS-active substrates

The overall SERS response, which typically results from a combination of EM and CM effects, is largely dependent on the substrate performance. In rough noble metal (e.g. Au, Ag, and Cu) nanostructures with high free carrier densities, the SERS signal amplification is largely attributed to their strong LSPR effects in the visible-to-near infrared range (Fig. 3d) [56, 57]. Furthermore, by precisely tuning the composition, size, morphology and surrounding medium of plasmonic nanomaterials, the EF can reach or exceed 10¹² [58]. Despite the fascinating SERS activity, the fabrication of plasmonic nanostructures typically involves expensive or hazardous precursors and complex synthetic processes, which result in poor reproducibility and instability, collectively hindering large-scale production and applications. Moreover, the inherent photothermal effect of plasmonic nanostructures may alter or damage biological samples [59].

On the contrary, non-metallic nanostructures (e.g., inorganic semiconductors, carbon-based materials and organic polymers) have distinctive features, including superior cost-effectiveness, high chemical stability, large-scale uniform fabrication capability, biocompatibility, and tunable physicochemical properties, making them ideal alternative candidates for SERS detection (Fig. 3e) [60, 61]. Nevertheless, their SERS activity is governed primarily by CM effect, with the enhancement typically limited to a relatively low EF (10³-10⁵). To date, chemically stable and uniform non-metallic nanostructures with excellent synergistic enhancement of EM and CM activities have been developed through heterojunction construction, morphology design, phase regulation, and defect engineering [62–65]. For instance, Jiang’s group [66] proposed flower-like MoS₂ microspheres with abundant sulfur vacancies, whose EF could reach up to 2.54 × 10⁸ due to the synergistic EM and CM effects.

Analogously, hybrid metallic/non-metallic SERS composite substrates can effectively integrate the EM and CM effects, thereby yielding superior SERS performance for highly sensitive and rapid bioanalysis [67, 68]. Gong et al. [69] presented Au nanorods/Ti₃C₂T_X/nickel foams (AuNRs/MXene/NFs) nanocomposite with an EF of up to 10⁷ via synergistic effects, demonstrating their application in rapid and sensitive detection of quetiapine in urine with a low detection limit. Collectively, the hybrid SERS composite substrates not only facilitate mechanistic investigations on synergistic SERS signal amplification but also guide the rational design and fabrication of highly sensitive substrates for bioanalysis. However, it remains challenging to quantitatively determine the contributions of the EM and CM effects and to construct a clear mechanistic explanation model for such hybrid systems.

SERS detection strategies

At present, SERS detection methodologies can be classified into label-free and labeled strategies, which involve the direct acquisition of the analyte SERS spectra and the indirect assessment of analyte content variations on the basis of specific SERS nanotags, respectively.

Label-free SERS strategy for direct analysis

In the label-free SERS strategy, detailed molecular vibrational fingerprints of biomarkers can be directly acquired, facilitating early cancer diagnosis without tedious sample preparation [70–72]. While this convenient strategy enables rapid detection of diverse targets, its performance is limited not only by the accessibility and affinity between analytes and substrates, but also by the inherent Raman scattering cross-sections of the analytes. Furthermore, the overlapping Raman signals from complex biological matrices, poor signal reproducibility under dynamic physiological conditions, background noise interference, and intricate analyte‒substrate interactions collectively pose challenges for reliable quantitative analysis. Therefore, research efforts have been devoted to fabricating promising SERS-active substrates with enhanced sensitivity and specificity, while chemometrics and AI algorithms are utilized to decipher complex SERS spectra effectively [73, 74].

Labeled SERS strategy for indirect analysis

Unlike the label-free strategy, the labeled SERS strategy employs exogenous Raman probes to achieve reliable multiplexed quantitative determination in complex matrices. Generally, labeled SERS signals originate from well-designed nanotags, which are typically constructed by embedding plasmonic nanoparticles and Raman reporter molecules into a protective shell, followed by functionalization with specific recognition and targeting moieties. In a typical labeled SERS assay, noble-metal nanostructures with tailored compositions and morphologies are selected to amplify Raman signals. Meanwhile, Raman reporter molecules with sharp, narrow, and clearly distinguishable characteristic peaks are chosen to facilitate high-throughput and multiplexed SERS sensing. Lastly, protective shells modified with specific recognition and targeting moieties are designed to stabilize the SERS nanotags, improve their biocompatibility, and confer specific targeting ability.

The prominent advantage of the labeled SERS strategy lies in not only providing exceptional target specificity but also enabling high-throughput multiplex detection in complex environments via distinct Raman reporter-coded nanotags [75]. For instance, Yu et al. [76] employed various Raman reporters (MGITC, DTDC, RBITC) to design labeled SERS nanotags for the identification and quantification of influenza A and COVID-19, whereas Shim et al. [77] adopted 2,3,5,6-tetrafluoro-4-mercaptobenzoic acid (TFMBA) and 4-mercaptobenzoic acid (4-MBA) to construct labeled SERS nanotags for the ultrasensitive and simultaneous detection of Alzheimer’s disease-related biomarkers Aβ42 and Aβ40. However, the fabrication of labeled SERS nanotags is time-consuming and labor-intensive, and unable to provide SERS fingerprint spectra of analytes. Table 2 summarizes the analytical performance of label-free and labeled SERS strategies. Consequently, appropriate SERS strategies for early-stage UGI cancer diagnosis should be selected after carefully evaluating the distinctive advantages of label-free and labeled SERS approaches.

Table 2.

Performance comparison between label-free and labeled SERS strategies

Aspects	Label-free strategies	Labeled strategies
Targets	Metabolic molecules, biological fluid samples, cancer cells, EVs, etc.	Cancer-associated biomarkers (VOCs, exosomes, ctDNA, miRNA, glutathione, etc.)
Advantages	Direct rapid detection, detailed molecular vibrational fingerprints, convenient operability, nondestructive	High-throughput, exceptional sensitivity, excellent specificity, quantitative analysis, rapid response, nondestructive
Limitations	Poor signal reproducibility, difficulties in data interpretation, intricate analyte–substrate interactions, severe impurity interferences	Complex fabrication, high costs, reliance on extrinsic Raman reporters, time consumption, inevitable nonspecific adsorption, batch-to-batch variations
Clinical Readiness level	Validated primarily with cultured cells and biofluid samples, with moderately sized retrospective patient cohorts	Validated with small- or moderate-scale retrospective patient cohorts, results compared with those of standard detection methods

Integrated SERS-based strategies for UGI cancer diagnosis

AI-assisted label-free SERS strategy

AI-assisted SERS platforms for Gastric Cancer (GC) diagnosis

AI is a domain of computer science that aims to mimic human intelligence behaviors through algorithms and data-driven approaches [78]. Generally, the label-free SERS approach enables the direct and highly sensitive detection of intrinsic vibrational spectral differences in the biological fluids or tissues of patients with UGI cancers [14]. However, this strategy is restricted by the complex biological detection environment, leading to overlapping SERS spectral features, signal variability, and subsequent difficulties in data interpretation [79]. AI algorithms can extract underlying features from complex spectral datasets, decipher overlapping Raman signals, and identify subtle molecular fingerprint variations, holding great potential for advancing the label-free SERS approach in the diagnosis of early-stage UGI cancers.

GC is a prevalent malignancy characterized by high mortality and poor early detection, representing a critical threat to human health. Currently, various AI algorithms, including principal component analysis (PCA), linear discriminant analysis (LDA), multiple local means-based nearest neighbor (MLMNN), centroid displacement-based nearest neighbor (CDNN), and the support vector machine (SVM), have been applied in label-free SERS diagnostic strategy for the early diagnosis of GC [44]. For example, Guo et al. [80] employed PCA and a novel characteristic ratio method (CRM) to distinguish the Raman spectral differences between GC patients and healthy individuals, yielding high sensitivity and specificity in reflecting abnormal metabolite fluctuations in GC patients’ serum. Moreover, PCA, as an effective dimensionality reduction technique, can also be integrated with other machine learning models to improve overall classification performance [81]. Specifically, Wu and coworkers [82] fabricated a spiral inertial microfluidic chip for cell separation and enrichment, and developed a PCA-LDA model for SERS analysis, which was ultimately applied to discriminate human gastric cancer (HGC) cells from similarly sized nonmalignant gastric epithelial (GES-1) cells and white blood cells (WBCs) (Fig. 4a). Obviously, this strategy yielded exceptional classification accuracy of 92.7% and a specificity of 97.8%, underscoring its potential for early-stage GC detection (Fig. 4b).

Fig. 4.

Machine learning-assisted SERS detection for early GC diagnosis. a) Schematic of PCA-LDA applied to SERS analysis for distinguishing human GC cells; b) PCA-LDA-based SERS analysis for discrimination and identification of HGC, WBC, and GES-1 cells. Reproduced with permission [82]. Copyright 2025, American Chemical Society. c) Schematic of SERS analysis integrated with PCA-CDNN model to distinguish serum samples from mice with precancerous lesions. Reproduced with permission [86]. Copyright 2024, Springer. d) Schematic illustration of LDA-SVM-based SERS analysis for discriminating sEVs from GC and non-GC participants; e) tracking analysis of the unique exosome fingerprints from GC patients. Reproduced with permission [87]. Copyright 2022, by the authors under CC-BY licence

Additionally, the MLMNN method, a high-performance classifier improved from traditional k-nearest neighbors, generates class-specific representative local mean vectors and performs classification via distance metrics to these prototypes [83]. Subsequently, Chen et al. [84] developed a novel PCA-MLMNN model to identify critical serum spectral features from patients with gastric lesions of different grades, achieving outstanding classification performance, with 96.7% sensitivity and 95% specificity. Apart from MLMNN, the CDNN model is another powerful discriminatory model based on the nearest neighbor classification concept [85]. Cao et al. [86] fabricated an agaric-shaped Au nanoarray substrate to capture SERS signals from mouse serum at different gastric lesion stages (precancerous lesions and GC stage). According to SERS spectral analysis, the elevated characteristic peaks at 667, 941, 1439, 1540, and 1721 cm^{− 1} were attributed to the abnormal metabolism of collagen, polysaccharides, lipids, proteins and steroids. In contrast, the continuously declining peak at 1618 cm⁻¹ corresponded to the substantial depletion of peripheral tryptophan in the tumor microenvironment. These distinct spectral features allowed the PCA-CDNN algorithm to accurately identify the gastric precancerous lesion group. Ultimately, this federated methodology delivered exceptional classification performance, achieving up to 100% in detection accuracy, sensitivity, and specificity (Fig. 4c).

Small EVs (sEVs), which are pivotal for intercellular communication, carry parental cell-specific molecular information and serve as valuable biomarkers for early UGI cancer diagnosis. However, the coexistence of normal and cancer-derived sEV subtypes in clinical samples frequently causes misclassification, impairing diagnostic accuracy. Therefore, Liu et al. [87] established a relabeling strategy to analyze the biochemical compositions of individual sEVs based on their single‑vesicle SERS fingerprints. By extracting GC-specific sEV subtypes via spectral cross-comparison, the SVM‑based classification (with LDA for dimensionality reduction and visualization) achieved highly accurate discrimination between GC and noncancer samples (n = 15 per group) (Fig. 4d). Notably, this approach successfully identified nine patient-specific sEV types across tissue, blood, and saliva, laying the groundwork for research into vesicle biogenesis (Fig. 4e).

Furthermore, deep learning models, such as artificial neural network (ANN), convolutional neural networks (CNNs), offer superior nonlinear fitting capacity, robust data processing, and autonomous feature extraction, and have been applied in label-free SERS diagnosis. For example, Xie et al. [88] trained an ANN model on 1780 SERS spectra collected from exhaled breath samples of healthy individuals (H, n = 49), lung cancer patients (LC, n = 22) and GC patients (n = 8), delivering high classification sensitivity and robust predictive performance (Fig. 5a). This study revealed elevated hydrocarbon levels (peak at ~ 1190 cm^{− 1}) in the exhaled breath of LC patients and increased aldehyde/ketone concentrations (peak at ~ 1441 cm^{− 1}) in GC patients compared with healthy individuals. These subtle differences in volatile organic compounds (VOCs) offer a noninvasive and reliable approach for cancer diagnosis (Fig. 5b). Furthermore, using SERS spectral datasets collected from 210 healthy controls and 543 cancer patients (including 60 GC patients), Shin et al. [89] integrated a multiple instance learning paradigm with CNN-based classifiers to identify early-stage malignancies. Ultimately, this model achieved robust stratification for early-stage cancers classified as stage II or earlier according to the American Joint Committee on Cancer (AJCC) Tumor–Node–Metastasis (TNM) staging system (Fig. 5c). Additionally, one-dimensional SERS spectral data can be converted into two-dimensional images to improve the visualization of early cancer-related spectral signatures and satisfy the input requirements of CNN architectures. Specifically, Lin et al. [90] analyzed serum SERS spectral data obtained from 1896 healthy controls and 1655 early-stage cancer patients (including 100 GC and 38 EC cases), identifying spectral differences that distinguished the cancer cohort from controls. These predominant differential Raman peaks located at 454, 494, 638, 1012, 1134 and 1662 cm⁻¹ corresponded to altered amino acid metabolism, perturbed monosaccharide metabolism and abnormal protein conformational structures, all associated with tumorigenesis. Subsequently, the authors converted one-dimensional SERS spectra into two-dimensional images via heatmap transformation and continuous wavelet transform for dimensionality enhancement, while preserving both global and local features (Fig. 5d). Ultimately, an 18-layer residual neural network (ResNet18) was employed for spectral image classification, which achieved excellent diagnostic accuracy in discriminating early-stage cancer patients (AJCC stage 0, I, and II) from healthy controls, thus indicating its potential for clinical cancer diagnosis (Fig. 5e).

Fig. 5.

Deep learning-assisted SERS detection for early GC diagnosis. a) Application of an ANN model to the SERS breath sample dataset for classification and outcome prediction; b) schematic of exhaled VOC variations based on average normalized SERS spectra from the H, LC, and GC groups. Reproduced with permission [88]. Copyright 2024, Elsevier. c) Workflow of cancerous exosome source prediction using CNN-based SERS data analysis. Reproduced with permission [89]. Copyright 2023, Springer Nature. d) Two-dimensional image data obtained by dimension enhancement and e) classification performance evaluation of the ResNet18 model. Reproduced with permission [76]. Copyright 2025, Springer Nature

AI-assisted SERS platforms for Esophageal Cancer (EC) Diagnosis

EC is frequently associated with a high mortality rate and poor prognosis; therefore, the development of effective early-stage diagnostic approaches is essential for improving patient outcomes [15, 91]. Notably, AI-assisted SERS platforms also exhibit considerable potential for early-stage EC diagnosis, facilitating the implementation of personalized therapeutic strategies [92–95]. For example, Zhang et al. [96] employed a PCA-LDA algorithm to process the SERS spectral datasets obtained from two human esophageal tumor cell lines (KYSE and TE) and one normal oral epithelial cell line (Het-1 A) on a CoSe₂/CoS₂ heterojunction substrate. SERS measurements revealed distinct spectral variations among the three cell lines at 1350, 1480, and 1600 cm⁻¹. Typically, the peaks at 1350 and 1480 cm⁻¹ corresponded to nucleic acids, collagen, and lipid chains, while the band at 1600 cm⁻¹ originated from aromatic amino acids. Such spectral disparities were attributed to abnormal DNA replication, disrupted lipid metabolism and increased protein levels in tumor cells. Consequently, the PCA-LDA model enabled reliable discrimination between normal Het-1 A cells and two malignant esophageal cell lines (KYSE and TE), yielding favorable sensitivity and specificity (Fig. 6a).

Fig. 6.

AI algorithms-assisted SERS detection for early EC diagnosis. a) Diagram of SERS-active substrate synthesis and PCA-LDA model application for discriminating EC cells (KYSE and TE) from normal cells (Het-1 A). Reproduced with permission [96]. Copyright 2025, Wiley. b) Workflow of sample pretreatment process and PLS-DA model applied for distinguishing EC patients from healthy controls based on the collected mean SERS spectra. Reproduced with permission [97]. Copyright 2017, Elsevier

Moreover, supervised partial least squares discriminant analysis (PLS-DA) offers an effective alternative for the multi-class classification of SERS spectra, as it explicitly extracts latent components that maximize inter-group separability. Based on its inherent features, Feng and coworkers [97] employed the PLS-DA algorithm to discriminate the SERS fingerprint profiles of urinary modified nucleosides from healthy individuals (n = 52), nasopharyngeal cancer patients (n = 62) and EC patients (n = 55), achieving high sensitivity (90.9%) and specificity (98.2%) for EC diagnosis (Fig. 6b). Furthermore, Liu et al. [98] indicated that the orthogonal partial least squares discriminant analysis (OPLS-DA) enabled clear differentiation between EC cells and their secreted exosomes, as well as accurate discrimination of exosomes originating from different EC cell lines, thus verifying its high predictive value for EC diagnosis.

Collectively, AI-assisted SERS platforms enable the extraction of critical features from complex spectral datasets, facilitating the identification of subtle spectral variations and the development of rapid and convenient diagnostic approaches for early-stage UGI cancers. A summary of representative AI-assisted SERS platforms is presented in Table 3, while the performance of the corresponding algorithms is compared in Table 4. Apparently, it is essential to choose an appropriate AI algorithm tailored to the specific research objectives. However, AI algorithms remain underexplored in the early diagnosis of EC, with most current studies relying on conventional machine learning methods limited to extracting basic spectral features. Therefore, it is necessary to develop innovative AI models to advance early-stage EC diagnosis. Besides, systematic and transparent interpretation of its “black-box” mechanism is critical to accelerate the clinical translation of AI-assisted SERS platforms.

Table 3.

AI-assisted SERS platforms for early diagnosis of UGI cancers

AI algorithms	SERS substrates	Detection targets	Biological samples	Detection performance	Ref.
GC diagnosis
CRM/PCA	Au NPs	Nucleic acids, amino acids, carbohydrates	Human serum	Sensitivity: CRM (100%) PCA (94.1%). Specificity: CRM (97.4%) PCA (94.1%) Accuracy: CRM (98.5%) PCA (94.1%)	[80]
PCA-LDA	Ag NPs	Human GC cells, gastric epithelial cells, and white blood cells	Simulated gastric juice, simulated ascites	Accuracy: 96% Specificity: 100%	[82]
PCA-MLMNN	Au lotus-shaped nanoarrays	Serum samples from patients with different grades of gastric lesions	Human serum	Sensitivity: >96.7% Specificity: >95.0% Accuracy: 97.5%	[84]
PCA-CDNN	Agaric-shaped Au nanoarray substrate	Biomarkers, including proteins, lipids, carbohydrates, and amino acids	Mouse serum	Accuracy: 100% Sensitivity: 100% Specificity: 100%	[86]
LDA-SVM	Au nanopyramid array	sEVs	Human tissue, blood, saliva	The accuracies of identifying GC versus healthy controls were 90% (tissue), 85% (blood), and 72% (saliva), respectively.	[87]
ANN	Plasmonic metal organic frameworks nanoparticle film	Glutaraldehyde	Human exhalation	Accuracy: 89%	[88]
CNN	Au nanoparticle-aggregated array chip	Plasma exosomes	Human plasma	Sensitivity: 90.2% Specificity: 94.4% Accuracy: >95%	[89]
ResNet18	Ag NPs	Metabolic molecules in the serum	Human serum	Accuracy: 94.75%	[90]
EC diagnosis
PCA-LDA	Crystalline-amorphous CoSe2/CoS2 heterojunction	Esophageal-related tumor cells (KYSE, TE), healthy oral epithelial cells (Het-1 A)	Clone cells	Accuracy of KYSE: 91.67% TE: 90.32% Het-1 A: 92.86%	[96]
PLS-DA	Au NPs	Modified nucleosides in urine	Human urine	Sensitivity: 90.9% Specificity: 98.2% Accuracy: 95.8%	[97]
OPLS-DA	Au NPs	EC cells and their exosomes	Cell lines and their exosomes	Accuracy: 100%	[98]

Table 4.

Comprehensive performance comparison of the employed algorithms

Algorithm types	Advantages	Limitations	Sample size requirements	Clinical implementation capacity
CRM	Extremely low computational complexity, fully interpretable decision logic	Restricted to binary classification tasks, losing global feature information, high sensitivity to random noise	≥ 10 samples per class	Large-scale preliminary screening tool
PCA	Unsupervised linear dimensionality reduction, visualization of sample cluster distribution, label-free modeling, no overfitting risk as it is a deterministic projection	High sensitivity to extreme outliers and interfering samples, only supports linear dimensionality reduction	Sample-to-variable ratio exceeded 5:1	Universal preprocessing and quality control tool for clinical high-dimensional data
PCA-LDA	Low computational resource demand, fast training speed, interpretable discriminant functions	Information loss, only supports linear decision boundaries, high sensitivity to unbalanced class distribution	≥ 20 samples per class, depending on the retained PCA dimensions	Linearly separable clinical datasets
PCA-MLMNN	Stronger robustness to local noise and blurred class boundaries than standard KNN, no overfitting risk, good reproducibility	Extremely low computational efficiency for large-scale clinical datasets, only supports linear distance measurement, No feature weight learning mechanism	15 ~ 50 samples per class, requires samples-per-class> dimension after PCA	Auxiliary classification model for single-center small-sample clinical cohorts
PCA-CDNN	Stronger robustness to abnormal outlier samples, simple model logic, stable performance in small-sample clinical cohorts with blurred class boundaries	Low efficiency for large-volume clinical batch detection, poor discrimination ability for nonlinear dataset, No feature weight learning mechanism	15 ~ 50 samples per class, requires samples-per-class> dimension after PCA	Small-sample disease screening, classification scenarios with a small number of abnormal interfering samples
LDA-SVM	Reduces the computational cost and overfitting risk of SVM, good generalization performance in high-dimensional small-sample datasets, supports kernel function replacement	The computational complexity of kernel SVM scales quadratically/cubically with sample size, making it inefficient for very large cohorts, multi-classification tasks need cumbersome modeling process	Sample-to-feature ratio for each class in the training set exceeded 5:1	Single-center medium-sample complex disease differential diagnosis
PLS-DA	Simple calculation process, interpretability, excellent adaptability to high-dimensional spectral data	Weak anti-interference ability, poor separation effect for complex nonlinear samples, false positive diagnosis under extremely unbalanced class distribution, highly prone to overfitting if the number of latent variables is not rigorously optimized via cross-validation, model performance is unstable with very small samples	≥ 15 ~ 20 samples per class	Classic baseline modeling tool, single-center small-sample clinical cohort disease diagnosis
OPLS-DA	Stronger denoising ability, interpretability, allows objective assessment of overfitting via permutation tests and cross-validation	Poor separation effect for complex nonlinear samples, multi-class disease staging/subtyping need cumbersome modeling process	≥ 15 ~ 20 samples per class	Powerful diagnostic tool for spectral disease diagnosis, multi-center clinical validation of diagnostic models
ANN	Simple network structure, one-dimensional spectral feature analysis, low computing resource demand	No local feature extraction capability, high susceptibility to overfitting with limited training samples, dependence on strong regularization strategies to learn complex nonlinear representations	≥ 100 samples per class, typically requires dimensionality reduction (e.g., via PCA) and heavy data augmentation	Auxiliary classification model for one-dimensional spectral data
CNN	Automatically extracts local spatial features, lower overfitting risk than ANN, strong nonlinear fitting ability	Requires more clinical cases, limited interpretability (black-box nature)	≥ 300 samples per class	One-dimensional Raman identification, auxiliary screening model
ResNet18	Superior classification performance and strong discriminative capacity, pretrained weights available for transfer learning, strong robustness	High requirements for large clinical cohorts and high-performance GPUs, higher overfitting risk when trained from scratch, poor interpretability, complex model structure	≥ 500 samples per class (training from scratch)	High-precision AI diagnostic model with critical constraints on computational resources, cohort scale and model interpretability

Labeled SERS strategies for UGI cancer diagnosis

While the label-free SERS strategy can directly distinguish spectral differences in metabolic fluctuations between UGI cancer patients and healthy controls, its effectiveness is constrained by poor reproducibility and instability, which are attributed to nonspecific adsorption and random distribution of hotspots during detection. To overcome these limitations, numerous labeled SERS-integrated strategies have been widely developed, including probe molecule-based, functional nucleic acid (FNA)-based, microfluidic-based, clustered regularly interspaced short palindromic repeats (CRISPR)-based, and endoscopy-integrated SERS strategies.

Probe molecule-based SERS diagnostic strategy

Typically, Raman reporters with distinct and intense spectral signals can be employed as intermediaries to trace target molecules. Specifically, the presence of target analytes induces characteristic signal changes in Raman reporters (peak shifts, spectral shape alterations, and intensity variations), effectively converting elusive target signals into readily detectable reporter responses and thus greatly extending the application of SERS in biomedicine (Table 5).

Table 5.

Probe molecule-based SERS strategies for early diagnosis of UGI cancers

Raman reporters	SERS substrates	Mechanisms	Detection targets	Biological samples	Detection performance	Ref.
4-MBA	Au nanosheets	Target biomarker binding induced nanomechanical stress	Protein biomarkers (CEA, VEGF)	Serum	Limit of detection (LOD) of CEA: 0.38 pg/mL VEGF: 0.82 pg/mL Results consistent with ELISA	[99]
4-ATP	Ag@ZIF-67	MOF shell enriched VOCs; Schiff base reaction	Aldehyde/Ketone VOCs	Exhaled breath	Discrimination accuracy: 89.83% Sensitivity: 91.23% Specificity: 88.52% Area under curve (AUC): 0.9715	[100]
4-ATP	S-CNF/Au NSs flexible substrate	Schiff base reaction	MDA	Serum	LOD: 1.29 × 10− 11 M Results consistent with colorimetric method	[101]
4-ATP	Au@CsPbBr3	EM effects, CM effects, Schiff base reaction	GC-related aldehydes	Exhaled breath	LOD: 3.29 × 10− 9 M Discrimination accuracy: 81.09%	[102]
4-ATP	Au NSs	Sandwich structure, Breakage of β-glycosidic bond β-lactose	SA-β-gal	Serum	LOD: 4.79 × 10− 5 U/mL Results consistent with colorimetric method	[103]
DTNB	Ag NPs-BCM	GSH cleaved the disulfide bond of DTNB	GSH	Serum	LOD: 2.9 × 10− 8 M	[104]
3-MPBA	Co-TA@Ag NPs	Peroxidase-mimicking activity, electron transfer, and hydroxyl alcohol esterification reaction	Chiral amino acids (D-Pro, D-Ala)	Saliva	LOD: D-Pro: 2.37 µM D-Ala: 2.15 µM Validation with actual samples	[105]

For example, Huang et al. [99] fabricated a 4-MBA modified Au nanosheet array (GNS/4-MBA) for antibody immobilization. In the presence of two serum tumor biomarkers, carcinoembryonic antigen (CEA) and vascular endothelial growth factor (VEGF), the GNS/4-MBA/antibody conjugate underwent nanomechanical deformation accompanied by Raman frequency shifts at ~ 1593 cm⁻¹, facilitating GC diagnosis. Moreover, a tubular SERS sensor based on 4-aminothiophenol (4-ATP)-labeled ZIF-67-coated silver particles (Ag@ZIF-67) was developed to detect various aldehyde and ketone VOCs (GC breath biomarkers) via spectral changes induced by the Schiff base reaction [100]. Similarly, Wang et al. [101] employed 4-ATP to functionalize flexible 2D nanocellulose-Au nanostars (S-CNF/Au NSs) substrate for the detection of malondialdehyde (MDA), which is overexpressed in GC patient serum (20 clinical serum samples). The MDA content was quantified based on the intensity ratio (I_C=N/ I_1078cm−1), where I_C=N refers to the peak from the Schiff base reaction product between 4-ATP and MDA, and I_1078cm−1 is the characteristic peak of 4-ATP as an internal reference (Fig. 7a). This approach enabled the highly sensitive and selective detection of MDA, thereby effectively differentiating GC patients from controls (Fig. 7b).

Fig. 7.

Probe molecule-based SERS approaches for ultrasensitive metabolic biomarker analysis. a) Schematic of MDA detection based on SERS signal changes of 4-ATP-modified 2D S-CNF/Au NSs substrate; b) assessment of detection performance and GC diagnostic capabilities. Reproduced with permission [101]. Copyright 2024, Springer Nature. c) The SERS spectral variations of 4-ATP functionalized Au@CsPbBr₃ film for identifying GC-related aldehydes; d) the selective discrimination capability for different aldehydes and comprehensive classification assessment between GC breath samples and healthy control samples. Reproduced with permission [102]. Copyright 2022, Wiley. e) Schematic for quantitative determination of GSH in human serum via the SERS spectral variations of DTNB; f) the corresponding substrate morphological characterization and detection feasibility analysis. Reproduced with permission [104]. Copyright 2023, Springer Nature. g) Schematic diagram of the fabricated 3-MPBA-functionalized Co-TA@AgNP composite for the detection of D-Pro and D-Ala in GC patients’ saliva; h) SERS detection performance analysis. Reproduced with permission [105]. Copyright 2025, Elsevier

Furthermore, Man et al. [102] constructed a 4-ATP-functionalized perovskite-molecule charge transfer complex (Au@CsPbBr₃ film) and clearly distinguished 9 types of GC-related aldehydes via machine learning-assisted analysis (Fig. 7c). The clinical utility of this approach was validated by its high diagnostic accuracy (81.09%) in differentiating breath samples of GC patients from those of healthy controls (18 GC patients and 12 healthy individuals) (Fig. 7d). Notably, 4-ATP-labeled SERS tags have also been constructed to determine the enzymatic activity of β-galactosidase (SA-β-gal), contributing to GC diagnosis. Specifically, SA-β-gal cleaved the β-galactoside bond in β-lactose, which triggered the release of SERS tags and produced a distinct signal reduction, enabling highly sensitive and selective discrimination between serum samples from GC patients and healthy controls [103].

Beyond 4-ATP, other reporter molecules have been employed to monitor the fluctuations of UGI-associated biomarkers. For instance, Li et al. [104] utilized 3D flexible AgNP-decorated bacterial cellulose membrane (AgNPs-BCM) and 5,5’-dithiobis (2-nitrobenzoic acid) (DTNB) molecules to detect serum glutathione (GSH) levels in a cohort of 30 subjects (Fig. 7e). GSH selectively cleaved the disulfide bond of DTNB, resulting in distinct spectral changes that enabled the quantitative determination of serum GSH levels for sensitive GC diagnosis (Fig. 7f). Additionally, Yao’s group [105] developed 3-mercaptophenylboronic acid (3-MPBA)-functionalized bimetallic heterostructure nanocomposites (Co-TA@AgNPs) with dual chiral recognition capability and peroxidase-mimicking activity, which achieved sensitive detection of salivary D-amino acids (D-proline/D-Pro and D-alanine/D-Ala), allowing for noninvasive early-stage GC diagnosis (Fig. 7g). By monitoring the intensity ratio between the unique Raman peak at 654 cm⁻¹ (attributed to the interactions between D-amino acids and 3-MPBA) and an internal standard peak at 1077 cm⁻¹, this method established a robust quantitative platform for D-amino acids (Fig. 7h).

Notably, monitoring specific changes in the Raman signals of probe molecules enables the sensitive detection of cancer-related biomarkers, thereby facilitating the early screening of UGI cancers. Therefore, future efforts should focus on developing novel probe molecules that are not only highly specific to different analytes but also undergo reliable and measurable signal changes upon target recognition.

FNA-based SERS diagnostic strategy

As a category of nucleic acid molecules with distinct structural specificity (e.g., hairpin DNA, tetrahedral DNA), FNAs have been regarded as ideal sensing components for high-performance biosensors due to their programmable molecular recognition, conformational flexibility, and multiplexing capability [106, 107]. FNA-based SERS sensors can not only generate abundant “hot spots” for remarkable signal amplification, but also possess specific targeting ability to enrich multiple analytes, thereby enabling sensitive, specific and multiplexed detection of various biomarkers [108, 109]. Therefore, numerous FNA-based SERS sensors have been proposed for the ultrasensitive and multiplex detection of UGI cancer-associated biomarkers, which is conducive to the early diagnosis and screening of UGI cancers (Table 6) [110].

Table 6.

FNA-based SERS strategies for early diagnosis of UGI cancers

FNAs	SERS substrates	Mechanisms	Detection targets	Biological samples	Detection performance	Ref.
ssDNA	MoS2NSs probes and AgNRs array electrodes	Sandwich structure, Specific capture ability	miR-106a	Human serum	LOD: SERS: 67.44 fM EC: 248.01 fM	[112]
ssDNA	Au trioctahedral nanoparticles arrays	Specific capture ability, Duplex-specific nuclease enzymatic cleavage	miR-21, miR-25	Human serum	LOD: miR-21: 4.29 aM miR-25: 8.12 aM Positive detection rate: 93.33%	[113]
hpDNA	Au rhombic dodecahedron	Competitive binding ability	miR-96-5p	Human serum	Results comparable to qRT–PCR Time: 30 min	[114]
hpDNA	Au NPAs and Au@Pt NRs	Peroxidase-mimicking activity, CHA amplification	miR-196b, miR-221	Human Serum	LOD: miR-196b: 2.21 aM miR-221: 3.09 aM, Results comparable to qRT–PCR	[115]
hpDNA	Au@Pt NPs, Au-coated Si nanopillar arrays	Peroxidase-mimicking activity CHA amplification	EFNA1 and MMP13	Human Serum	LOD: EFNA1: 0.75 pg/mL MMP13: 0.84 pg/mL Results comparable to ELISA	[116]
triApt-TDN	Ag NRs array and Au NPs	bHCR amplification	Exosomes	Clinical blood samples	LOD: 0.39 particles/µL, Time: 60 min, 2 µL sample consumption volume AUC: 0.987	[118]
MATD	Ag NRs assay and Au NPs	CHA amplification	Exosomes	Clinical blood samples	LOD: 2.98 × 103 particles/mL, Time: 40 min 2 µL sample consumption volume AUC: 1	[119]

Complementary single-stranded DNA (ssDNA) can be rationally designed to specifically recognize and capture microRNA (miRNA) targets, which serve as effective biomarkers for UGI cancer screening [111]. Significantly, Zhai et al. [112] constructed a novel SERS-electrochemical (EC) dual-mode sensor for the ultrasensitive and specific detection of GC-associated miR-106a in human serum, by immobilizing ssDNA P1 on Ag NRs array electrodes and labeling ssDNA P2 onto multi-functionalized MoS₂ (mF-MoS₂ NSs) probes (Fig. 8a). The presence of miR-106a could be specifically recognized and captured by ssDNA P1 and ssDNA P2, which triggered the assembly of a sandwich structure (mF-MoS₂ NSs probes/miR-106a/AgNR array electrodes), yielding both detectable square wave voltammetry and SERS signals (Fig. 8b). Similarly, Chen et al. [113] engineered ssDNA-functionalized SERS nanoprobes to realize the sensitive and specific detection of miR-21 and miR-25 (overexpressed in gastric adenocarcinoma patients) with analytical performance comparable to that of traditional quantitative real-time polymerase chain reaction (qRT‒PCR).

Fig. 8.

FNA-based SERS strategies for UGI cancer-associated miRNAs and exosomes detection. a) Schematic of the designed EC-SERS dual-mode biosensor for miR-106a detection; b) the corresponding SERS spectra and EC signals. Reproduced with permission [112]. Copyright 2022, Elsevier. c) Schematic illustration of the hpDNA-mediated SERS sensing platform for the sensitive detection of GC-related miRNAs (miR-196b and miR-221); d) the corresponding SERS detection sensitivity and method validation. Reproduced with permission [115]. Copyright 2025, Elsevier. e) Design of an integrated SERS sensor coupling TDN-based trivalent aptamer with bHCR for detection of GC cell-derived exosomes; f) evaluation of sensing performance and clinical detection accuracy. Reproduced with permission [118]. Copyright 2025, Elsevier. g) Scheme of MATD-mediated SERS strategy for quantitative detection of GC cell-secreted exosomes; h) assessment of sensing performance and clinical utility of this sensor. Reproduced with permission [119]. Copyright 2025, Elsevier

Hairpin DNA (hpDNA), featuring a characteristic double helical stem and a compact loop, has also been employed as a key sensing element in FNA-based SERS platforms. For example, Jia et al. [114] proposed an hpDNA-mediated SERS competitive lateral flow bioassay for sensitive detection of GC-associated miR-96-5p, which is driven by the high binding affinity between hpDNA-biorecognition probes and target miR-96-5p. Furthermore, hpDNA-mediated SERS platforms can be integrated with catalytic hairpin assembly (CHA) to achieve the efficient detection of GC-related biomarkers (miR-196b and miR-221) in 90 clinical serum samples [115]. Typically, hpDNA2-fixed Au nano pumpkin arrays (Au NPAs@HP2) were prepared as SERS-active substrates, while hpDNA1-modified platinum-coated Au nanorods (Au@Pt NRs@HP1) acted as nanozyme probes to catalyze the TMB-H₂O₂ reaction, yielding oxTMB as SERS signal reporter with distinct peaks (Fig. 8c). In the presence of target miRNAs, the opening HP1 initiated the CHA cycle due to the strong complementary effect between HP1 and HP2, drawing Au@Pt NRs to the surface of Au NPAs and generating a remarkably amplified SERS signal for ultrasensitive detection of gastric precancerous lesions (Fig. 8d). Analogously, Dai and coworkers [116] extended this integrated strategy to the sensitive and specific detection of early GC-associated protein biomarkers (Ephrin-A1/EFNA1 and Matrix Metalloproteinase 13/MMP13) via the design of antibody‒DNA conjugates, achieving LODs down to the pg/mL level within a 40-minute assay time. With verified diagnostic performance in 20 early-stage GC patients and 20 healthy controls, this sensing platform exhibited great promise for rapid, highly sensitive, and high-throughput screening of early GC biomarkers.

Three-dimensional tetrahedral DNA nanostructures (TDNs) with excellent structural rigidity, superior stability, and high programmability have also been exploited to design SERS biosensors for early UGI cancer screening [117]. For instance, Liu et al. [118] combined TDN-based trivalent aptamers (triApt-TDNs) with branched hybridization chain reaction (bHCR) to fabricate a SERS biosensor toward the ultrasensitive detection of GC cell-associated exosomes (total n = 20). Firstly, triApt-TDNs were immobilized onto the AgNR array for the specific recognition of MUC1 proteins, a biomarker overexpressed on the surface of GC cell-derived exosomes. Subsequently, trigger aptamers (tgApts) anchored on exosomes initiated bHCR via specific binding with hpDNA1 (H1)-functionalized SERS tags, generating branched DNA concatemers and driving intense SERS signal enhancement (Fig. 8e). Overall, the triApt-TDN-mediated SERS sensor enabled highly specific, reproducible, and molecular-level detection of target exosomes, facilitating reliable discrimination between GC patients and healthy controls (Fig. 8f). Furthermore, Zhang et al. [119] merged multivalent aptamer-linked TDN (MATD) with a CHA signal amplification strategy to develop a promising SERS sensor for the sensitive detection of GC cell-secreted exosomes. In contrast to the aforementioned research, the MATD-functionalized AgNRs array specifically captured CD63, a ubiquitous surface marker of exosomes. Then the captured exosomes were recognized by the identification SERS tags via aptamer-MUC1 protein binding, triggering the release of patch strands to initiate a CHA cycle and induce the self-assembly of AuNP networks with abundant SERS “hot spots” (Fig. 8g). The corresponding superior SERS performance enabled the specific and ultrasensitive detection of GC cell-secreted exosomes, thereby achieving effective differentiation between clinical GC patients and healthy controls using a pilot clinical set of 20 samples (including 15 GC patients and 5 healthy controls) (Fig. 8h).

These outstanding performance features confirm that FNA-based SERS sensors exhibit high specificity, ultrahigh sensitivity, and multiplexing capacity for the quantitative detection of low-abundance UGI cancer-associated miRNAs and exosomes, underscoring their considerable potential for the early diagnosis of UGI cancers. Additionally, the FNA-based SERS sensing strategy holds substantial promise for integration with portable diagnostic platforms, paving the way for its translation into point-of-care diagnostic applications.

Microfluidic-based SERS diagnostic strategy

Microfluidic chips, commonly referred to as lab-on-a-chip (LoC), organ-on-a-chip, or micro-physiological systems, are miniaturized analytical platforms that consolidate various laboratory processes onto a single chip [120]. Owing to their minimal sample consumption, multifunctionality, customizability, and precise microscale fluid control capability, microfluidic chips have been applied in diverse fields such as chemical synthesis, biomedical analysis and environmental monitoring [121, 122]. In particular, the microfluidic-based SERS strategies combine the precise fluid manipulation capability of microfluidics with the high sensitivity and rapid response of the SERS technique, enabling the robust and accurate bioanalysis of UGI cancer biomarkers for early diagnosis (Table 7) [123].

Table 7.

Microfluidic-based SERS strategies for early diagnosis of UGI cancers

Microfluidic chip properties	SERS substrates	Mechanisms	Detection targets	Biological samples	Detection performance	Ref.
Self-driven capillary chip, hydrophilicity	Au nano-hexagonal arrays, Au nanobipyramid @Ag shell	Stronger binding between aptamer and target protein	MMP-9 and IL-6	Human Serum	LOD: MMP-9: 0.263 pg/mL IL-6: 0.195 pg/mL Time: 20 min Results consistent with ELISA	[124]
Capillary pump effects with satisfactory fluidity, magnetic adsorption effect	Fe3O4@AuNPs and AuNCs	Stronger affinity between aptamer and target protein, Magnetic-induced aggregation effects	MMP-9 and IL-6	Human Serum	LOD: MMP-9: 0.178 pg/mL IL-6: 0.165 pg/mL Time: 15 min Results comparable to ELISA and pathological findings	[125]
Capillary action, magnetic adsorption effect, hydrophilicity	Sesame seed globular Fe3O4@AuNPs	Enzymatic reaction, Magnetic-induced aggregation effects	D-Pro and D-Ala	Human Saliva	LOD: D-Pro: 6.8 µM D-Ala: 8.3 µM Time: 30 min Results consistent with pathological findings	[126]
Polyethylene glycol for hydrophilic treatment, pump-free, six parallel-channels	Au nanobowl array and Cu2O octahedra	Cascade amplification strategy (CHA-HCR)	ctDNA	Mouse Serum	LOD: PIK3CA and E542K: 1.26 aM TP53: 2.04 aM Time: 13 min Results consistent with qRT–PCR	[127]
Capillary force, hydrophilicity	Au nanoflower and Ag@Au nanorods	Form “sandwich structure” based on specific recognition and capture ability	Thrombin and PDGF-B	Human serum	LOD: Thrombin: 3.62 fM PDGF-B: 2.86 fM Time: 30 min Results consistent with ELISA	[128]
Polyethylene glycol for hydrophilic treatment, capillary action, three-channels with six detection chambers	Au nano-bipyramids	Competitive adsorption mechanism	VEGF and PDGF-B	Human serum	LOD: VEGF: 0.342 pg/mL PDGF-B: 0.265 pg/mL Time: 20 min Results comparable to ELISA	[129]

Zhuang et al. [124] designed a comb-shaped, self-driven capillary microfluidic SERS (LoC–SERS) platform for the rapid identification and sensitive detection of GC biomarkers (metalloproteinase-9/MMP-9, interleukin-6/IL-6). Fabricated via soft lithography, this platform comprised three independently operating microchannel modules, each containing six components: an inlet, a sample mixing zone, two detection zones, a capillary pump, and an outlet (Fig. 9a). Such an integrated LoC–SERS platform exhibited favorable hydrophilicity and sealing performance, which enabled autonomous capillary-driven flow within the channels and ultimately allowed for the sensitive and specific detection of MMP-9 and IL-6 based on the high binding affinity of aptamers (Fig. 9b). Based on SERS detection of 90 clinical serum samples (30 healthy controls, 30 patients with gastric precancerous lesions, and 30 GC patients), this microfluidic platform enabled rapid, automated, and ultrasensitive identification of precancerous lesions, and the detection results were further validated against standard clinical diagnostic approaches (Fig. 9c).

Fig. 9.

Microfluidic-based SERS systems for simultaneous detection of multiple UGI cancer-associated biomarkers. a) An integrated LoC–SERS platform for the duplexed detection of protein biomarkers (MMP-9 and IL-6); b) evaluation of the flow characteristics, SERS sensing response and operational stability; c) detection performance assessments and results validated by ELISA assay. Reproduced with permission [124]. Copyright 2025, Elsevier. d) A magnetically driven, pump-free LoC–SERS device for performing D-amino acid (D-Pro/D-Ala) detection in patient saliva; e) evaluation of experimental feasibility and detection efficiency; f) SERS spectral profiles of D-Pro and D-Ala, and specific discrimination of saliva samples from GC patients and healthy controls. Reproduced with permission [126]. Copyright 2024, Elsevier. g) A six-channel, pump-free microfluidic–SERS system for GC-derived ctDNA detection; h) microfluidic chip structural images and corresponding performance assessments; i) comparison of serum ctDNA levels between tumor-bearing and healthy control mice. Reproduced with permission [127]. Copyright 2022, Springer Nature. j) Schematic of a three-channel LoC–SERS system for multiplexed detection of GC-related protein biomarkers; k) experimental feasibility assessments; l) evaluation of detection sensitivity and classification analysis of serum samples. Reproduced with permission [129]. Copyright 2023, Elsevier

Microfluidic-based SERS devices can be further integrated with a miniature magnet to improve SERS performance via the magnetic aggregation effect. For instance, utilizing the separation and enrichment capabilities of a miniature magnet embedded beneath the detection chamber, Huang and coworkers [125] developed a magnetic microfluidic–SERS system that achieved simultaneous, rapid and ultrasensitive detection of MMP-9 and IL-6 in human serum, as well as efficient discrimination between GC patients and healthy individuals. Furthermore, by leveraging the synergistic effects of enzymatic cascade reactions and magnetic aggregation, a capillary-driven LoC–SERS device was fabricated for the specific and quantitative detection of GC biomarkers (D-Pro and D-Ala) (Fig. 9d) [126]. Typically, N45 circular magnets were incorporated into the mold during UV lithography, and the resulting hydrophilic LoC–SERS device enabled rapid and sensitive detection of D-Pro and D-Ala in GC patients’ saliva, demonstrating its ability to discriminate GC patients from healthy controls (20 clinical saliva samples) (Fig. 9e–f).

Currently, highly integrated multi-channel microfluidic–SERS systems have undergone rapid advancement. For instance, Cao et al. [127] developed a portable high-throughput microfluidic–SERS system, which enabled ultrasensitive detection of GC-related ctDNA via a CHA-HCR cascade signal amplification strategy (Fig. 9g). This system featured a six-channel parallel design and relied solely on capillary action for automatic fluid transport (Fig. 9h). Specifically, the target ctDNA (PIK3CA E542K or TP53) initiated a CHA-mediated HCR, leading to the self-assembly of long nicked double-stranded DNA nanowires that recruited numerous SERS tags to the plasmonic substrate, consequently generating a boosted SERS signal due to the formation of dense “hot spots”. Ultimately, the designed microfluidic‒SERS system enabled the ultrasensitive detection of GC-associated ctDNA within 13 min, and its results strongly correlated with those of qRT‒PCR in a mouse model, highlighting its considerable potential for practical utility (Fig. 9i). Analogously, employing photolithography and wet etching, Zhu et al. [128] developed a six-channel, pump-free LoC–SERS microfluidic chip. This chip achieved robust detection of GC-related protein biomarkers with performance comparable to that of ELISA, owing to pronounced SERS signal enhancement induced by the formed “sandwich” structure. Additionally, Chen and coworkers [129] proposed a three-channel, self-driven LoC–SERS system for simultaneous multiplexed detection of protein biomarkers (VEGF and platelet-derived growth factor-B/PDGF-B) in serum (30 GC patients and 30 healthy controls) based on the competitive adsorption mechanism between target proteins and Raman reporter-labeled complementary DNA strands (Fig. 9j–k). The excellent sensing performance indicated that the LoC–SERS system enabled rapid, accurate, and efficient biomarker detection, offering a practical auxiliary tool for early UGI cancer screening (Fig. 9l).

Consequently, microfluidic-based SERS platforms efficiently integrate sample processing, target enrichment, and sensitive SERS detection into a portable chip, facilitating rapid, high-throughput, and accurate bioanalysis with minimal sample consumption. Furthermore, the standardized fluidic control and high-precision fabrication also afford excellent reproducibility and sensitivity, rendering this strategy highly promising for early noninvasive diagnosis of UGI cancers.

CRISPR-based SERS diagnostic strategy

The CRISPR system, an adaptive immune mechanism found in bacteria and archaea, employs a guide RNA (gRNA) to direct CRISPR-associated (Cas) effector proteins to target or cleave specific nucleotide sequences [130]. Currently, CRISPR/Cas systems (e.g., based on Cas9, Cas13 or Cas12) are renowned for their gene-editing capacity and have also been engineered as powerful disease diagnostic tools, owing to their high efficiency, single-base recognition specificity, and flexible programmability [131]. The rational combination of CRISPR/Cas systems with SERS assays enables the ultrasensitive and specific detection of UGI-associated biomarkers, thereby facilitating rapid early screening (Table 8).

Table 8.

CRISPR-based or endoscopy-integrated SERS strategies for early diagnosis of UGI cancers

Diagnostic strategies	SERS substrates	Mechanisms	Detection targets	Biological samples	Detection performance	Ref.
CRISPR-based SERS systems
CRISPR/Cas13a	Ag NRs	bHCR amplification, EM effects	miR-106a	Human serum	LOD: 8.55 aM Time: 60 min Results in agreement with qRT–PCR AUC: 0.996	[133]
CRISPR/Cas13a	Ag NRs, Au NPs	CHA amplification, EM effects	SGC-7901 cells -derived exosomes	50% Human serum	LOD: 1.26 × 102 particles/mL Time: 60 min 2 µL sample consumption volume	[134]
CRISPR/Cas13a	Ag NRs, Au NPs	Trans-cleavage activity, DNAzyme-driven DNA walker amplification, EM effects	exosomal miR-106a	Human serum	LOD: 6.1 × 10³ particles/mL Time: 80 min AUC: 0.985, Results consistent with qRT–PCR	[135]
Endoscopy-integrated SERS systems
Fiber optic-based Raman device	Au@SiO2@Raman probe	White-light endoscopy offered visual observation, SERS provided molecular analysis	Excised human pathological tissues	Pork tissue and Human tissue	Laser power of 42 mW with 300 ms integration times Sensitivity level: femtomolar Multiplex detection	[137]
Dual-modality white-light/Raman clinical grade endoscopic imaging system	Au@IR780 dye@SiO2@PEG2000	White-light endoscopy provided macroscopic tissue context, Raman imaging realized tumor detection	Premalignant lesions in the esophagus, stomach, and colorectal	Transgenic mouse and rat models	Raman endoscopy identified 0.5–1.0 mm lesions Sensitivity level: femtomolar	[138]

The CRISPR/Cas13a system is a valuable RNA-targeting platform that utilizes CRISPR RNA (crRNA) to guide the recognition of specific target RNAs. Subsequently, target RNA recognition activates the ribonuclease (RNase) activity of Cas13a, leading to cis-cleavage of the target and nonspecific trans-cleavage of bystander RNAs [132]. Based on this principle, Zhang et al. [133] developed a novel CRISPR/Cas13a-based SERS sensor to achieve ultrasensitive detection of GC-related nucleic acid biomarker miR-106a. In the presence of target miR-106a, the activated CRISPR/Cas13a system produced numerous reconstituted trigger fragments, which then initiated a bHCR, immobilizing abundant Raman reporter-labeled hairpin probes on the SERS-active substrate and generating remarkable signal amplification. The detection results of 48 clinical serum samples (12 healthy donors, 12 patients with atypical gastric mucosal hyperplasia, and 24 GC patients) were highly consistent with those obtained by the gold-standard qRT–PCR, demonstrating that this approach is a promising tool for early GC-related biomarker detection.

Moreover, the CRISPR/Cas13a-based SERS platform was extended to the sensitive detection of GC-associated exosomes. For example, Song’s group [134] combined CRISPR/Cas13a system’s trans-cleavage activity, SERS technique’s EM enhancement and CHA strategy’s signal amplification to achieve the highly sensitive and specific detection of exosomes derived from GC cells (SGC-7901). Upon binding to target exosomes, the designed hairpin aptamer (MUC-apt) activated the trans-cleavage activity of CRISPR/Cas13a. The cleavage products subsequently initiated a CHA reaction, capturing abundant SERS tags onto the Ag NR array and producing dramatic SERS signal amplification (Fig. 10a). With single-base specificity, rapid response, low sample consumption and excellent sensitivity, this CRISPR/Cas13a-based SERS strategy allowed for low abundance exosome detection, contributing to early-stage GC diagnosis (Fig. 10b). Recently, He et al. [135] innovatively integrated the trans-cleavage activity of CRISPR/Cas13a, the cascade amplification effect of DNA walker, and the EM enhancement of SERS, achieving highly sensitive and specific quantification of GC-derived exosomal miR-106a in serum samples from 20 GC patients and 10 healthy volunteers. This sensing platform delivered favorable reproducibility and stability, demonstrating robust clinical diagnostic performance. Despite their potential, CRISPR-based SERS platforms are still in their infancy. The integration of other CRISPR/Cas systems (e.g., Cas9, Cas12) into SERS-based sensing is a key research priority, which would further expand the diagnostic toolkit for UGI cancers.

Fig. 10.

The emerging integrated strategies for the early screening of UGI cancers. a) Schematic of the CRISPR/Cas13a-based SERS system for GC-related exosome detection; b) evaluation of the detection performance of this system. Reproduced with permission [134]. Copyright 2024, American Chemical Society. c) Schematic illustration of an endoscopy-integrated SERS platform for the detection of various precancerous lesions; d) characterization and performance evaluation of the prepared PEGylated SERRS-NPs; e) SERRS-based Raman imaging for early detection of smaller gastroesophageal junction lesions, with corresponding histological results. Reproduced with permission [138]. Copyright 2019, American Chemical Society

Endoscopy-integrated SERS diagnostic strategy

Endoscopy is a primary diagnostic modality for UGI tumors, enabling direct visualization and real-time assessment of lesion morphology [136]. Conventional white-light endoscopy, however, has significant limitations in detecting early-stage lesions, as its diagnostic accuracy is susceptible to interobserver variability, leading to a high false-negative rate and restricting its broader application. The endoscopy-integrated SERS strategy bridges macroscopic visualization with microscopic analysis, providing not only morphological visualization of lesions but also detailed molecular information from biomarker profiling. Therefore, this integrated strategy allows for dual-mode, real-time, in vivo analysis, concurrently assessing both morphological and molecular features.

For instance, Zavaleta et al. [137] employed a noncontact Raman endoscope for real-time, multiplexed molecular imaging in excised human tissue specimens, enabling the simultaneous quantification of up to 10 distinct tumor-targeting SERS nanoprobes. Although this versatile system permitted highly sensitive molecular analysis over a wide working distance and reduced the miss rate of conventional endoscopy, the lack of in vivo animal validation rendered its clinical feasibility unproven. Moreover, the localized administration method used in this study failed to ensure sufficient and homogeneous contact between SERS nanoprobes and mucosal lesions within the intricate in vivo environment, thereby hindering its clinical translation.

Furthermore, Kircher’s group [138] introduced a dual-modality endoscopic-Raman imaging system based on surface-enhanced resonance Raman scattering nanoparticles (SERRS-NPs), which successfully identified precancerous lesions in the esophagus, stomach, and intestines of mouse and rat models (Fig. 10c). The applied SERRS-NPs were composed of a Au core, the near-infrared dye IR780, a silica shell and methoxy-terminated polyethylene glycol (PEG) and exhibited femtomolar sensitivity and extraordinary signal specificity (Fig. 10d). By leveraging the enhanced permeability and retention (EPR) effect, this powerful Raman endoscopic system sensitively identified premalignant lesions at the gastroesophageal junction in mice, and provided real-time Raman spectral information and relevant anatomical context (Fig. 10e). Thus, this dual-modality augmented Raman endoscopy system enabled the highly sensitive detection of dysplastic lesions, facilitating early diagnosis and guiding precise therapeutic interventions. Collectively, the endoscopy-integrated SERS strategy bridges macroscopic imaging and molecular-level analysis, enabling a shift from morphological inspection to precise molecular profiling for early cancer diagnosis, which holds significant promise in precision medicine. However, several critical obstacles must be overcome for successful clinical translation, including nanoparticle toxicity concerns, high instrumentation costs, laser safety issues, and the complexity of spectral data interpretation.

Challenges and future perspectives

UGI cancers remain a major global health challenge because of their prolonged asymptomatic phase and high mortality rates. While endoscopic biopsy remains the well-established gold standard, its clinical utility is compromised by invasiveness, high healthcare costs, and stringent technical requirements for practitioners. Additionally, conventional serological biomarkers exhibit insufficient sensitivity and specificity for early-stage UGI cancer diagnosis; thus, developing novel diagnostic strategies for early screening is essential to improve public health outcomes. As a powerful spectroscopic technique, SERS has emerged as a promising diagnostic tool for the early screening of UGI cancers. Generally, AI-assisted label-free SERS platforms, featuring easy operation, rapid spectral acquisition, ultrahigh sensitivity and nondestructive detection, yield comprehensive molecular fingerprints, enabling high-throughput identification of differential metabolite signatures between tumor and healthy samples. In parallel, integrated labeled SERS strategies exhibit considerable advantages in the highly specific and sensitive detection of diverse UGI cancer-associated biomarkers, further advancing early screening. Table 9 provides a systematic comparison of the diagnostic performance of these innovative SERS-based strategies. Despite their remarkable diagnostic potential, several critical challenges remain. The primary obstacles and future research endeavors are outlined as follows:

Table 9.

Performance comparison among different SERS-based diagnostic strategies

Strategy categories	Sample types	Biomarkers and corresponding LODs	Assay time range	Clinical sample size	Diagnostic performance
AI-assisted SERS platforms	Various biofluid samples, exhaled breath, cells	Metabolic molecules, exosomes, cancer cells, VOCs (mainly direct detection, without LOD assessment)	The procedure mainly includes sample pretreatment, SERS substrate preparation, ~ 20 min of SERS spectral acquisition, and subsequent data analysis	Ranging predominantly from tens to hundreds of subjects	Excellent diagnostic performance (high sensitivity, specificity and accuracy), rapid detection
Probe molecule-based SERS strategies	Serum, exhaled breath, saliva	Protein biomarkers (pg/mL), VOCs (nM), GSH (nM), D-amino acids (µM)	30 ~ 60 min	20 ~ 30 samples	High sensitivity, rapid detection, satisfactory diagnostic accuracy, results consistent with gold standard
FNA-based SERS strategies	Serum	miRNAs (aM ~ fM), exosomes (102~103 particles/mL), protein biomarkers (pg/mL)	30 ~ 60 min	20 ~ 90 samples	Rapid analysis, low sample consumption, high sensitivity and specificity, results consistent with gold standard
Microfluidic-based SERS chips	Serum, saliva	Protein biomarkers (pg/mL or fM), D-amino acids (µM), ctDNA (aM)	13 ~ 30 min	20 ~ 90 samples	Rapid analysis, high sensitivity and specificity, results consistent with gold standard
CRISPR-based SERS systems	Serum	miRNA (aM), exosomes (102~103 particles/mL)	60 ~ 80 min	30 ~ 48 samples	Low sample consumption, high sensitivity and specificity, results consistent with gold standard
Endoscopy-integrated SERS systems	Human tissue (ex vivo), mouse model	Excised human pathological tissues (SERS signal reached femtomolar sensitivity), premalignant lesions regions in mouse model (0.5 ~ 1.0 mm lesions)	SERS nanotags preparation, sample pretreatment/intravenous injection, endoscopic observation, ~ 20 min of SERS spectral acquisition, and subsequent data analysis	Only in vitro detection/ Animal model validation	Molecular level information, a wide working distance, high collection speed, tiny lesions detection, high sensitivity

Expanding SERS substrate diversity

Typically, label-free SERS strategies prioritize substrates that can both enrich target analytes and enhance their intrinsic Raman signals, whereas labeled SERS assays prefer highly SERS-active nanomaterials to construct specific SERS nanotags. Although noble-metal nanomaterials are extensively employed in SERS assays, they remain restricted by limited reproducibility, insufficient specificity, and poor stability. The emergence of non-metallic nanomaterials, particularly engineered semiconductors, has significantly diversified SERS substrate types by virtue of their selective enhancement capability, high sensitivity, and improved stability. Nevertheless, the rational design and development of high-quality SERS-active substrates still demand further efforts.

Recent studies have demonstrated that two-dimensional MXene, which possesses superior hydrophilicity, large specific surface area, high-density electronic states and abundant surface terminations, enables efficient exosome enrichment and facilitates subsequent SERS analysis [139, 140]. Additionally, borophene, the lightest member of the Xene family, has an ultrahigh carrier mobility, intrinsic metallic conductivity, and numerous surface anchoring sites [141, 142]. Notably, Au-coated borophene has exhibited satisfactory SERS activity, underscoring its potential as an advanced substrate for clinical diagnostics [143]. Apart from MXenes, metal-organic frameworks (MOFs) have also stood out as promising SERS substrates, benefiting from their ultrahigh porosity, tailorable pore architectures, inherent molecular sieving function, and robust structural stability [144]. Most pristine MOFs deliver unsatisfactory SERS performance in the absence of EM effects; therefore, MOF@noble metal composites serve as the dominant substrates for bioanalytical detection [145]. As representative examples, Wang et al. exploited AuNC@zeolitic imidazolate framework-8 (ZIF-8) core-shell nanostructures for lung cancer volatile biomarker sensing, whereas Yu’s group fabricated anodic aluminum oxide@Ag cavity/Ag@ZIF-8 nanoparticle to realize accurate multiplex quantification of biomarkers in sweat and exhaled breath [146, 147]. Interestingly, MOFs are ideal precursors for constructing single-atom catalysts (SACs). Benefiting from maximized atomic utilization and superior catalytic performance, AgSACs derived from AgMOF can efficiently accelerate the redox formation of AuNPs, generating remarkable SERS enhancement signals [148]. Besides, SACs can be anchored onto 3D Ag aerogels, and the resulting substrate simultaneously delivers robust oxidase-like nanozyme activity and remarkable SERS activity [149]. Collectively, these innovative nanomaterials represent promising candidate substrates for SERS-based UGI cancer diagnosis.

Reproducibility and standardization

The complex biological environment not only affects the reproducibility and sensitivity of label-free SERS analysis but also can alter the structure of SERS nanotags, posing a major challenge to SERS detection. Although various pretreatment approaches (e.g., extraction, chromatography, protein precipitation, ionic strength regulation, serial dilution) have been proposed to mitigate matrix interference, standardized protocols tailored to diverse biological samples are still lacking. Thus, establishing standardized sample pretreatment protocols is a fundamental prerequisite for robust clinical SERS analysis.

Furthermore, stable and highly reproducible SERS signals are largely determined by substrate quality, instrumental calibration protocols, and standardized spectral preprocessing procedures. In general, classical AuNPs/AgNPs are easy to fabricate, but they suffer from poor stability and randomly distributed hotspots. Although various advanced preparation strategies, such as top-down lithographic patterning, template-assisted nanofabrication, and ordered self-assembly methods, have been employed to fabricate high-quality SERS substrates with uniform nanoarchitectures and highly consistent plasmonic hotspots, more facile, time-efficient and low-cost manufacturing approaches are still required to facilitate large-scale production [150]. Meanwhile, universal calibration strategies (including internal standards and multi-reference calibrations), rigorous environmental control protocols, and real-time signal quality control methods are indispensable for mitigating interlaboratory discrepancies and instrumental variations, ensuring reproducible and reliable SERS signal acquisition [151]. Given the biological matrix complexity, including autofluorescence and impurity interferences, diverse spectral preprocessing algorithms have been proposed to suppress spectral noise, improve spectral quality, and facilitate accurate analysis of SERS data [152]. Notably, discrepancies among distinct spectral preprocessing algorithms may impair the reproducibility and reliability of SERS analytical outcomes. Overall, universal standardized workflows and reference spectral databases for SERS must be developed, which are analogous to the MIQE guidelines for qPCR, as these are essential for eliminating clinical translational gaps.

Clinical validation

Although several SERS-based diagnostic platforms for UGI cancer diagnosis have achieved satisfactory validation in cell lines and mouse models, relevant research is still limited to proof-of-concept studies. Critically, the majority of existing studies utilize commercially available cell lines cultivated under standard monoculture conditions with simplified culture media. Such highly controllable experimental conditions are incapable of recapitulating the complex matrix interference in human physiological fluids and the heterogeneous characteristics of native tumor microenvironments. Hence, the signal stability, biosafety and diagnostic efficacy of SERS probes remain questionable in practical clinical applications. Furthermore, current animal model experiments often suffer from limited cohort sizes, inadequate systematic biosafety assessment, insufficient blinded sample verification, and a lack of relevant tests on human samples, which collectively impede the subsequent clinical validation. Despite these obstacles, SERS-based in vitro diagnostic platforms have demonstrated considerable translational potential for liquid biopsy-based early tumor screening. Specifically, clinical biofluid samples such as serum, urine, and sweat enable convenient sampling, which facilitates large-scale blinded cohort trials. Such cohort-based blinded assessment allows comprehensive, unbiased quantification of SERS diagnostic performance, with all diagnostic results benchmarked against clinical gold standards. This rigorous verification is conducive to accelerating regulatory approval and clinical translation. However, most of the reported studies lack statistically robust cohort datasets with stratification by tumor stage; future research should prioritize prospective multicenter trials aligned with the AJCC TNM staging framework to systematically evaluate the diagnostic performance of SERS-based platforms for early UGI tumors (Stages 0–II).

Perspectives on AI-assisted SERS platforms for UGI cancer diagnosis

Despite the promising diagnostic performance of AI-assisted SERS platforms, considerable limitations still hinder their clinical translation and regulatory approval. First, most existing investigations employ simple case-control designs lacking rigorous adjustment for clinical confounders, and stratified subgroups encompassing different disease stages and precancerous lesions are rarely established. Such methodological limitations reduce the clinical applicability of AI-assisted SERS strategies in real patient cohorts. Second, most available datasets were obtained from single-center retrospective cohorts with limited sample sizes and restricted population diversity. These data constraints inevitably result in poor generalizability and high overfitting risks for deep learning-based diagnostic models. Third, deep learning models are often constrained by their inherent black-box nature, which undermines the interpretability of model predictions derived from complex high-dimensional SERS features [153]. The obscure correlation between vibrational fingerprints and underlying biological information confuses both clinicians and regulators, thereby impeding regulatory approval. To tackle this bottleneck, explainable artificial intelligence (XAI) technologies have been developed to demystify deep learning decision-making [154, 155]. Broadly speaking, feature importance evaluation methods are widely adopted to guide global screening and subsequently extract core spectral signatures [156, 157]. In particular, SHapley Additive exPlanations (SHAP), a game-theory-driven attribution algorithm, quantifies feature contributions at the single-wavenumber and regional levels, clarifying the key Raman bands responsible for model classification [158]. Meanwhile, gradient-weighted class activation mapping (Grad-CAM) and analogous visualization methods yield class-activation heatmaps that enable intuitive localization of the critical spectral regions dominating deep learning-based diagnostic predictions [159, 160]. Although these techniques have markedly improved model transparency, many developments are still needed, particularly in terms of the prediction accuracy, full-spectrum interpretability, and elimination of medical data biases [161]. Furthermore, it is worth noting that SERS spectral datasets contain sensitive patient clinical information, rendering them vulnerable to privacy leakage during data transmission, dataset sharing, and model training. These broad ethical concerns and data compliance risks associated with AI-assisted SERS platforms warrant serious consideration [162].

Perspectives on microfluidic-based SERS systems for UGI cancer diagnosis

Current microfluidic-based SERS systems for UGI cancer diagnosis are often constrained by oversimplified structural designs, which result in limited detection efficiency, insufficient anti-interference capacity, and poor reproducibility. Consequently, there is an urgent need to develop sophisticated integrated systems that leverage advanced strategies, including three-dimensional alternating current electrokinetics, droplet microfluidics, automated control, and digital microfluidics, to further improve detection performance. Moreover, existing microfluidic-based SERS systems are largely confined to UGI-related protein biomarker detection, whereas other promising tumor biomarkers, such as EVs, nucleic acids, CTCs, and VOCs, remain underexplored [163]. Therefore, future research efforts should focus not only on the rational design and engineering of microfluidic–SERS platforms with superior sensitivity, rapid response, high throughput, and outstanding specificity, but also on achieving simultaneous quantification of diverse tumor biomarkers to realize accurate early diagnosis of UGI tumors. Notably, sophisticated high-performance microfluidic–SERS devices often entail higher fabrication costs; thus, a delicate balance must be struck between manufacturing expenditure and overall detection performance [164, 165].

Furthermore, microfluidic-based SERS platforms are highly promising for developing robust multimodal analytical frameworks, owing to their exceptional modular compatibility with various analytical techniques [166]. By integrating multidimensional disease-associated information, this approach is expected to provide a comprehensive and precise assessment of UGI cancer progression [167]. Additionally, leveraging AI-driven data analysis, microfluidic–SERS systems exhibit remarkable potential in directly tracking intrinsic SERS fingerprints of metabolites [168].

Perspectives on CRISPR-based SERS strategies for UGI cancer diagnosis

CRISPR/Cas systems are categorized into two primary classes, six main types and numerous subtypes, with Cas9, Cas12, and Cas13 serving as the predominant effector nucleases in bioanalysis [169]. Nevertheless, this review reveals that current CRISPR-based SERS strategies for UGI cancer diagnosis are largely confined to the Cas13a system and predominantly target GC-associated biomarkers. These constraints in both effector nuclease selection and diagnostic scope collectively highlight the primary directions for future investigations. Moreover, the stringent reaction conditions required by CRISPR/Cas systems not only increase operational complexity and experimental costs, but also restrict CRISPR–SERS platforms to laboratory-based research, thereby impeding their clinical translational potential [169, 170]. Given that endogenous matrix interferents in clinical biofluids readily deactivate Cas nucleases and degrade gRNAs, promising future research directions cover the rational engineering of interference-tolerant Cas variants, the chemical modification of gRNAs to reduce off-target cleavage, and the development of antifouling SERS substrates with superior performance [171]. Furthermore, considering low-abundance nucleic acid biomarkers in complex biofluid samples, engineering high-performance CRISPR-based SERS platforms endowed with exceptional detection sensitivity and ultrahigh target specificity is indispensable for robust quantitative monitoring [172]. Beyond that, the integration of CRISPR-based SERS platforms with portable microfluidic devices stands out as an attractive strategy, as it enables miniaturization and automation, thereby facilitating subsequent clinical translation and commercialization [173, 174].

Development of miniature SERS-based point-of-care testing (POCT) devices

Label-free SERS assays feature rapid analysis, rich molecular fingerprint information, and convenient operability, whereas labeled SERS strategies are valued for their multiplexed biomarker detection capability, ultrahigh sensitivity, and excellent specificity. Notably, the integration of label-free SERS assays into POCT devices holds promise for the sensitive screening and real-time monitoring of dynamic changes in the expression of UGI cancer biomarkers, yet this field remains in the early stages of development. Moreover, labeled SERS strategies can be combined with POCT devices, and numerous microfluidic-based SERS chips have been employed for the multiplexed detection of UGI biomarkers. However, these proposed chips are still constrained by simplistic designs, prolonged incubation periods, and nonnegligible nonspecific adsorption. Beyond microfluidic platforms, SERS systems can also be coupled with other POCT platforms, including portable Raman spectrometers, lateral flow assays (LFAs), smartphone-assisted systems and so on. Such integrated sensing strategies reduce the reliance on cumbersome benchtop Raman spectrometers, offering compelling application potential validated in numerous analytical detection scenarios [175, 176]. Specifically, Guan et al. achieved rapid and sensitive on-site detection of pseudovirus-SARS-CoV-2 and its variants in simulated oropharyngeal swab samples within 5 min by employing a handheld Raman spectrometer (SHINS-P700T, 785 nm) [177]. Lu’s group fabricated a self-calibrating SERS-LFA strip with high reproducibility, easy operation and low cost, enabling sensitive and accurate detection of prostate cancer biomarker [178]. Taking full advantage of the high portability, wide accessibility and user-friendly operation of smartphones, Zhang and coworkers proposed a smartphone-based portable Raman analyzer equipped with a wireless communication system, which enabled real-time SERS signal acquisition and wireless remote spectral processing [179]. Collectively, these advanced SERS-integrated POCT platforms have substantial potential for rapid, robust and accurate UGI diagnosis in resource-constrained clinical scenarios.

Development of SERS-based dual/multimodal diagnostic platforms

In early UGI cancer screening, the developed SERS-based diagnostic approaches are often plagued by shallow penetration depth, matrix interference, and inadequate disease-related information. The development of SERS-based dual-/multimodal diagnostic platforms integrates the synergistic strengths of SERS and other cutting-edge technologies, effectively mitigating the inherent limitations of standalone SERS. Typically, multifunctional magnetically actuated microrobots enabled drilling-mediated minimally invasive tissue sampling, facilitating targeted biopsy of deep-seated tracheal microlesions through synergistic integration with SERS biosensing [180]. Furthermore, the combination of SERS and molecular imprinting techniques enhanced the selective enrichment of target molecules in complex biofluid matrices, while the designed SERS/fluorescence dual-modal bio-probes yielded complementary information by fusing visual fluorescence imaging with detailed SERS spectra [181, 182].

Development of SERS-integrated multiomics analytical platforms

A global paradigm shift is underway in modern medicine, transitioning from “one-size-fits-all” approaches to precision medicine, namely, individualized strategies developed on the basis of a molecular understanding of personalized disease phenotypes [183]. Multiomics integration acts as a pivotal driving force underlying this transformative medical trend. It facilitates the high-throughput profiling of multi-class biomolecular signatures, allowing the comprehensive analysis of disease pathogenesis and the elucidation of the intricate underlying molecular networks [184, 185]. Benefiting from ultrahigh sensitivity, rapid spectral acquisition, unique molecular fingerprint information and nondestructive characteristics, SERS has evolved into a powerful platform for metabolic profiling. To date, numerous studies have demonstrated its prominent advantages in identifying disease-specific metabolic signatures and facilitating dynamic metabolic analysis [186–188]. Nevertheless, SERS-based metabonomics still encounters considerable challenges in accurate quantification of trace metabolites, reliable molecular spectral assignment, complex spectral data processing, and comprehensive coverage of endogenous metabolic molecules [189]. Furthermore, SERS can be synergistically combined with conventional metabolomics and proteomics for the in-depth elucidation of disease mechanisms [190]. Yet, the integration of SERS with genomics and transcriptomics remains largely underexplored. Moving forward, researchers should focus on the development of advanced SERS-integrated multiomics approaches to realize personalized, precise early diagnosis of UGI cancers.

Development of SERS imaging-based multifunctional platforms

Currently, endoscopy-integrated SERS strategies remain in their infancy, with relatively few studies reported in this field. Future efforts should concentrate on developing novel SERS-based imaging platforms capable of visualizing both subtle cancerous tissue alterations and the spatial distribution of biomarker abundance, which is essential for early UGI cancer detection and precise tumor localization. Furthermore, SERS imaging can be coupled with diverse therapeutic technologies, achieving simultaneous diagnosis and treatment of UGI cancers. This theranostic strategy has been validated in other cancer models [191]. Specifically, Liu et al. [192] utilized SERS/photoacoustic imaging to guide photothermal/photodynamic therapy, allowing for precise diagnosis and efficient ablation of breast tumors. Collectively, SERS imaging-based multifunctional platforms hold considerable potential for constructing integrated theranostic systems for UGI cancers, yet their biocompatibility, stability, detection depth, imaging speed, spatial resolution, and sensitivity require comprehensive evaluation prior to clinical translation.

Conclusion

In summary, this review systematically discusses recent progress in integrated SERS-based platforms for early UGI cancer diagnosis. Following a critical evaluation of routine diagnostic modalities, we underscore the superior performance of SERS-based strategies for the noninvasive, rapid, highly sensitive, and specific detection of metabolic alterations in patients, which are pivotal for early screening of UGI malignancies. Specifically, this review summarizes AI-assisted label-free SERS diagnostic approaches and promising labeled SERS strategies, the latter encompassing probe molecule-based, FNA-based, CRISPR-based, and endoscopy-integrated SERS platforms. While these integrated SERS-based strategies exhibit considerable potential, several nonnegligible challenges persist and impede their clinical translation. To address these challenges, we highlight key directions for future research, including expanding the diversity of SERS substrates, establishing standardized clinical protocols, developing universal calibration methodologies, and constructing versatile, high-performance SERS-based diagnostic platforms. Eventually, through the collaborative endeavors of scientific researchers, medical practitioners, and social support systems, coupled with the continuous advancement of interpretable AI models, SERS-based diagnostic strategies are expected to transition from attractive laboratory studies to standardized clinical screening programs, thus realizing their potential as next-generation precision diagnostic tools.

Abbreviations

UGI cancers

Upper gastrointestinal cancers

SERS

Surface-enhanced Raman spectroscopy

IARC

International Agency for Research on Cancer

MRI

Magnetic resonance imaging

CT

Computed tomography

PCR

Polymerase chain reaction

ELISA

Enzyme-linked immunosorbent assay

AI

Artificial intelligence

PTR-ToF-MS

Proton-transfer-reaction time-of-flight mass spectrometry

GC-MS

Gas chromatography-mass spectrometry

LC-MS

Liquid chromatography-mass spectrometry

PLSS

Polarized light scattering spectroscopy

CLE

Confocal laser endomicroscopy

CTCs

Circulating tumor cells

ctDNA

Circulating tumor DNA

EVs

Extracellular vesicles

NGS

Next-generation sequencing

EM

Electromagnetic enhancement mechanism

CM

Chemical enhancement mechanism

LSPR

Localized surface plasmon resonance

CT

Charge transfer

EF

Enhancement factor

TFMBA

2,3,5,6-tetrafluoro-4-mercaptobenzoic acid

4-MBA

4-mercaptobenzoic acid

GC

Gastric cancer

PCA

Principal Component Analysis

LDA

Linear Discriminant Analysis

MLMNN

Multiple Local Means-based Nearest Neighbor

CDNN

Centroid Displacement-based Nearest Neighbor

SVM

Support Vector Machine

CRM

characteristic ratio method

HGC

human gastric cancer cells

GES-1

nonmalignant gastric epithelial cells

WBCs

white blood cells

sEVs

small extracellular vesicles

ANN

Artificial Neural Network

AJCC

American Joint Committee on Cancer

TNM

Tumor Node Metastasis

CNNs

Convolutional Neural Networks

VOCs

volatile organic compounds

ResNet18

18-layer residual neural network

EC

esophageal cancer

PLS-DA

Partial Least Squares Discriminant Analysis

OPLS-DA

Orthogonal Partial Least Squares Discriminant Analysis

FNA

functional nucleic acid, CRISPR, clustered regularly interspaced short palindromic repeats

GNS

Au nanosheet array

CEA

carcinoembryonic antigen

VEGF

vascular endothelial growth factor

4-ATP

4-aminothiophenol

MDA

malondialdehyde

SA-β-gal

β-galactosidase

DTNB

5,5’-dithiobis (2-nitrobenzoic acid)

GSH

glutathione

3-MPBA

3-mercaptophenylboronic acid

D-Pro

D-proline

D-Ala

D-alanine

ssDNA

Complementary single-stranded DNA

miRNA

microRNA

EC

electrochemical

qRT-PCR

quantitative real-time polymerase chain reaction

hpDNA

Hairpin DNA

CHA

catalytic hairpin assembly

EFNA1

Ephrin-A1

MMP13

Matrix Metalloproteinase 13

TDNs

tetrahedral DNA nanostructures

triApt-TDNs

TDN-based trivalent aptamers

bHCR

branched hybridization chain reaction

tgApts

trigger aptamers

MATD

multivalent aptamer-linked TDN

MMP-9

metalloproteinase-9

IL-6

interleukin-6

ctDNA

circulating tumor DNA

PDGF-B

Platelet-derived growth factor-B

LoC

Lab-on-a-chip

EPR

Enhanced permeability and retention

MIQE

Minimum information for publication of quantitative real-time PCR experiments

POCT

Point-of-care testing

XAI

Explainable artificial intelligence

Grad-CAM

Gradient-weighted class activation mapping

SHAP

SHapley Additive exPlanations

SACs

Single-atom catalysts

MOF

Metal-organic framework

ZIF-8

Zeolitic imidazolate framework-8

SARS-CoV-2

Syndrome coronavirus 2

Ti₃C₂T_X

MXene

gRNAs

Guide RNAs

LFAs

Lateral flow assays

Author contributions

Q. W.: Investigation, Writing the original draft, Picture collection, Review and editing. S. J.: Writing, Organizing documents, Picture processing. S. Z.: Investigation, Data curation, Formal analysis. W. Y.: Investigation, Data curation, Picture processing. Y. S.: Supervision, Conceptualization, Editing. H. Z.: Project administration, Review and editing, Supervision. R. Z.: Supervision, Validation, Funding acquisition.

Funding

This work was supported by the National Natural Science Foundation of China (U24A6012), Shanxi Province Science Foundation for Youths (Grant 202303021222146), the National Key R&D Program of China (2023YFC3402800).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

All authors have provided consent for the manuscript to be published in Journal of Nanobiotechnology.

Competing interests

The authors declare no competing interests.