Artificial intelligence and radiomics in small cell lung cancer: a narrative review of imaging and multisource data integration

Abstract

Background and Objective

Artificial intelligence (AI) and radiomics are increasingly used to analyze imaging and multisource data in small cell lung cancer (SCLC), but the directness and clinical maturity of the available evidence remain uncertain. This narrative review aimed to summarize and critically evaluate SCLC-specific, mixed-cohort, and indirect evidence across diagnostic classification, prognostic prediction, treatment-associated outcomes, and lesion segmentation.

Methods

A structured literature search was conducted to identify human studies evaluating AI, deep learning (DL), radiomics, or machine learning within imaging-based or imaging-integrated workflows. Studies were selected according to clinical relevance, directness of SCLC evidence, methodological strength, and reporting completeness, and were classified as SCLC-specific, mixed-cohort, or indirect/transferable evidence.

Key Content and Findings

Direct SCLC-specific evidence was strongest for multicenter computed tomography (CT)-based prognostic models and lesion segmentation, whereas diagnostic classification relied mainly on mixed cohorts or public image-level datasets. Treatment-response models showed preliminary value but were predominantly retrospective and lacked prospective clinical validation. Retrospective studies suggested associations between model-derived risk scores and outcomes following prophylactic cranial irradiation (PCI), but did not establish treatment-benefit prediction. Pathomic and radiogenomic analyses also identified associations between quantitative features, immune-cell infiltration, genomic alterations, and treatment response; these findings remain hypothesis-generating. Interpretation is limited by small SCLC cohorts, retrospective single-center designs, non-independent image-level analyses, and limited external validation and interpretability.

Conclusions

AI and radiomics show potential in SCLC, particularly for CT-based prognosis and segmentation. However, treatment-response prediction, non-invasive subtype classification, and clinical decision support remain investigational and require prospective multicenter validation and patient-level external evaluation before clinical implementation.

Introduction

Small cell lung cancer (SCLC) is the most aggressive and prognostically poorest subtype of lung cancer, accounting for approximately 14% of all lung cancer cases (1). Its unique biological behavior is characterized by rapid proliferation, early widespread metastasis, and initial treatment sensitivity with a high risk of recurrence (2). In clinical practice, precise diagnosis and stratified treatment of SCLC face numerous challenges. SCLC overlaps radiologically with some non-small cell lung cancer subtypes, leading to limited accuracy of conventional diagnostic methods in differential diagnosis. Moreover, SCLC exhibits substantial interpatient heterogeneity, with marked variations in response to chemotherapy, radiotherapy, and immunotherapy, and there is a lack of effective biomarkers to guide individualized treatment decisions (2,3).

SCLC remains a clinically urgent disease despite recent therapeutic progress. Approximately 70% of patients present with extensive-stage disease at diagnosis, and even with modern systemic therapy, durable disease control is achieved only in a minority of patients (4). Current treatment paradigms increasingly combine platinum-etoposide chemotherapy with immune checkpoint inhibition in extensive-stage disease, while concurrent chemoradiotherapy followed by consolidation immunotherapy has become an important direction for limited-stage disease (4-8). In the IMpower133 trial, the addition of atezolizumab to carboplatin and etoposide significantly improved survival compared with chemotherapy alone, establishing chemoimmunotherapy as a first-line standard for extensive-stage SCLC (5,6). Similarly, the CASPIAN and ADRIATIC studies support the expanding role of programmed death ligand-1 (PD-L1) blockade in extensive-stage and limited-stage clinical settings, respectively (7). Nevertheless, the magnitude of benefit varies substantially across individuals, and many patients experience rapid relapse after an initial response (4-7). These clinical characteristics highlight the need for non-invasive, repeatable, and biologically informative tools capable of capturing tumor burden, metastatic distribution, treatment sensitivity, and dynamic disease evolution.

From an imaging perspective, SCLC is particularly suitable for artificial intelligence (AI)-based imaging and data-integration research because clinical decisions depend heavily on baseline staging, response evaluation, and surveillance of intrathoracic and intracranial disease. Computed tomography (CT) provides anatomical information regarding primary tumors, nodal involvement, and thoracic treatment fields; positron emission tomography/CT (PET/CT) adds metabolic characterization and may improve detection of occult metastatic disease; magnetic resonance imaging (MRI) is indispensable for assessing brain metastases and intracranial treatment planning. However, conventional visual interpretation is limited in quantifying intratumoral heterogeneity, peritumoral changes, and subtle longitudinal variations that precede RECIST-defined progression. Radiomics and deep learning (DL) can transform standard-of-care images into high-dimensional quantitative biomarkers, thereby providing a potential bridge between routine clinical imaging and individualized decision-making (9,10). Therefore, the value of DL in SCLC should not be restricted to binary classification; rather, it should be evaluated according to whether it can improve the full clinical pathway, including diagnosis, staging, risk stratification, treatment selection, response monitoring, and recurrence prediction.

AI-based image analysis includes conventional radiomics/machine-learning pipelines using predefined features and DL architectures, such as convolutional neural networks and Transformers, that learn task-relevant representations directly from data (11). Here, multimodal imaging is reserved for combinations of radiological imaging modalities, such as CT, PET, and MRI, whereas combinations of radiological and digital pathology images are termed radiology-pathology image integration. Models combining imaging with clinical variables, serum biomarkers, genomic data, or liquid-biopsy measurements are described as imaging–non-imaging data integration. These integration strategies can be implemented using either conventional machine learning or DL (11). Applications of these techniques in lung cancer have expanded from simple lesion detection and segmentation to higher-level clinical decision-support tasks including molecular subtype prediction, prognosis assessment, and treatment response stratification.

This narrative review, supported by a structured literature search, summarizes research on AI and radiomics using multimodal imaging and imaging-non-imaging data integration for the precision diagnosis and stratified treatment of SCLC. We distinguish DL, conventional radiomics/machine learning, multimodal imaging, radiology-pathology image integration, and imaging-non-imaging data integration across four clinical tasks: diagnostic classification, prognosis prediction, treatment-response prediction and treatment-associated outcomes, and lesion segmentation. SCLC-specific studies constitute the primary evidence base. Mixed lung cancer studies are included only when they report SCLC-specific results separately or provide directly relevant methodological evidence; studies without separately reported SCLC results are treated as indirect or potentially transferable evidence. The literature search, eligibility criteria, evidence classification, and study-selection procedures are described below. We present this article in accordance with the Narrative Review reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1693/rc).

Methods

Review design and literature search

This article was designed as a narrative review supported by a structured literature search, rather than a systematic or scoping review. PubMed/MEDLINE, Web of Science Core Collection, and Scopus were searched from database inception through May 31, 2026. The search was last conducted on June 2, 2026. Google Scholar and the reference lists of relevant reviews and eligible articles were also screened to identify additional studies. The search combined disease-related terms (‘small cell lung cancer’, ‘SCLC’, and ‘lung cancer’) with method-related terms (‘deep learning’, ‘artificial intelligence’, ‘machine learning’, ‘radiomics’, and ‘pathomics’) and data-source-related terms (‘multimodal imaging’, ‘data integration’, ‘computed tomography’, ‘CT’, ‘PET/CT’, ‘MRI’, ‘pathology’, ‘whole-slide image’, ‘clinical data’, ‘serum biomarker’, ‘liquid biopsy’, and ‘genomics’). MeSH terms and corresponding free-text terms were adapted to each database and combined using Boolean operators. English-language and human-study filters were applied, without a study-design filter. The literature-search strategy is summarized in Table 1, and the complete PubMed search strategy is provided in Table S1.

Table 1. Summary of the literature search strategy.

Items	Specification
Date of search	June 2, 2026 (last search)
Databases and other sources searched	PubMed/MEDLINE, Web of Science Core Collection, and Scopus; Google Scholar and the reference lists of relevant reviews and eligible articles were used for supplementary searching
Search terms used	Disease terms: ‘small cell lung cancer’, ‘SCLC’, and ‘lung cancer’; method terms: ‘deep learning’, ‘artificial intelligence’, ‘machine learning’, ‘radiomics’, and ‘pathomics’; data-source terms: ‘multimodal imaging’, ‘data integration’, ‘computed tomography’, ‘CT’, ‘PET/CT’, ‘MRI’, ‘pathology’, ‘whole-slide image’, ‘clinical data’, ‘serum biomarker’, ‘liquid biopsy’, and ‘genomics’. MeSH terms and corresponding free-text terms were combined with Boolean operators and adapted to each database. English-language and human-study filters were applied; no study-design filter was used. The detailed PubMed strategy is provided in Table S1
Timeframe	From database inception to May 31, 2026
Inclusion and exclusion criteria	Included: English-language, peer-reviewed, full-text human studies evaluating deep learning, radiomics, or machine learning in an imaging-based or imaging-integrated workflow relevant to diagnostic classification, prognosis/survival, treatment response or treatment-associated outcomes, or lesion segmentation. SCLC-only studies and mixed lung-cancer studies with separately identifiable SCLC results were treated as direct evidence; other relevant lung-cancer studies were retained only as indirect/transferable methodological evidence
	Excluded: animal or in vitro studies, duplicate publications, studies unrelated to lung cancer, and studies with insufficient methodological or outcome information. Conference papers or preprints were retained only for directly relevant emerging methods and were treated as preliminary evidence.
Selection process	Two authors (Y.Y. and J.S.) independently screened titles/abstracts and then full texts against the eligibility criteria. Disagreements were resolved through discussion; unresolved conflicts were adjudicated by a third senior author (J.D.). Representative studies were prioritized according to directness for SCLC, clinical relevance, methodological strength, validation design, and completeness of reporting

SCLC, small cell lung cancer.

Eligibility, evidence classification, and selection of representative studies

Original human studies were eligible for the evidence synthesis when they evaluated DL or an associated radiomics or machine-learning component within an imaging-based or imaging-integrated workflow and addressed at least one of the four clinical tasks considered in this review: diagnostic classification, prognosis or survival prediction, treatment-response prediction and treatment-associated outcomes, and lesion segmentation. SCLC-only studies were classified as direct SCLC evidence. Mixed lung cancer studies were classified as direct evidence only when SCLC-specific results were reported separately; studies without separately reported SCLC results were retained only as indirect or potentially transferable methodological evidence and were not used to support SCLC-specific clinical conclusions. Reviews, guidelines, and landmark clinical trials were used only to provide clinical or methodological background. Animal or in vitro studies, duplicate publications, studies unrelated to lung cancer, and studies with insufficient methodological or outcome information were excluded. Peer-reviewed full-text articles were prioritized; conference papers or preprints were retained only when they described an emerging method directly relevant to the review and were treated as preliminary evidence.

Two authors (Y.Y. and J.S.) independently screened retrieved records at the title/abstract and full-text levels against these criteria. Disagreements were resolved through discussion; unresolved conflicts were adjudicated by a third senior author (J.D.). Studies were selected for detailed discussion and inclusion in Table 2 according to the following criteria: directness of evidence for SCLC; relevance to one of the four clinical tasks; representativeness or novelty of the imaging source, non-imaging data source, and integration strategy; methodological strength, including cohort size, multicenter design, and internal or external validation; and completeness of performance and clinical reporting. When multiple studies addressed the same clinical task using similar methods, priority was given to SCLC-specific, multicenter, externally validated, and more completely reported studies. Owing to substantial heterogeneity in study populations, data modalities, clinical endpoints, and performance measures, the findings were synthesized narratively, and no meta-analysis was performed. Reported performance was interpreted according to the total sample size, number and proportion of SCLC cases, class distribution, unit of analysis (patient, scan, slide, or image), validation design, dataset source, and reported uncertainty. Results based on image-level cross-validation or internal random splitting were not considered equivalent to independent patient-level external validation. Confidence intervals (CIs) were reported when available.

Table 2. Representative AI and radiomics studies categorized by clinical endpoint, cohort characteristics, validation, and relevance to SCLC.

Clinical endpoint/task	Representative study	Imaging source(s) and additional non-imaging data	Analytical method and fusion strategy	Main performance	Clinical implication	Current limitation	Cohort, design, and validation	Evidence relevance to SCLC
Differential diagnosis between SCLC and NSCLC brain metastases	Sui et al.	Contrast-enhanced brain MRI	DL: ResNet18 MRI classifier	Accuracy: 75.2%	Demonstrates the feasibility of using MRI-based deep learning to distinguish SCLC from NSCLC in patients with brain metastases	Moderate diagnostic accuracy; limited sample size; requires external validation	n=192; SCLC =73 (38.0%); mixed SCLC/NSCLC cohort; retrospective; center setting NR; patient-level five-fold cross-validation; no independent external validation	Mixed-cohort evidence; 73 SCLC cases included
Lung cancer subtype classification	Jacob et al.	PET and CT images (multimodal imaging)	DL: attention-based CNN ensemble	Accuracy: 98.8% for three-class classification	Mixed-cohort image-level classification; patient-level SCLC validation was not reported	Public dataset setting may not fully reflect real-world SCLC heterogeneity	Patient n NR; 3,212 image slices, including 1,141 SCLC slices (35.5%); mixed public datasets; image-level ten-fold cross-validation; no patient-level external validation	Mixed-cohort, image-level evidence; patient-level SCLC results NR
Cell-level pathological classification	Tian et al.	Lung cancer cell images	DL: FPN-SE-ResNet18 with multiscale features	Accuracy: 98.84%	Potentially transferable cell-image classification method; no SCLC cases were included	Cell image datasets differ from routine whole-slide pathological workflows	750 public histopathology images; SCLC =0; NSCLC dataset; image-level five-fold cross-validation; no independent external validation	Indirect/transferable evidence; no SCLC cases included
Peripheral SCLC versus lung adenocarcinoma	Wang et al.	CT images/radiomics and radiographic features; clinical variables (non-imaging)	Conventional radiomics/ML with imaging-clinical integration: CT radiomics with LASSO and logistic regression	AUC: 0.967, higher than clinical-only and radiomics-only models	Supports the value of integrating clinical and imaging phenotypes for difficult peripheral lung cancer differentiation	Retrospective design; focused on a specific differential diagnosis scenario	n=240; SCLC =80 (33.3%); mixed SCLC/adenocarcinoma cohort; retrospective, single-center; random 7:3 internal split (training =170, validation =70); patient-level; no external validation	Mixed-cohort evidence; 80 SCLC cases included
Multi-class lung cancer pathological classification	Peng et al.	PET and CT images/radiomics; clinical variables (non-imaging)	Conventional radiomics/ML with imaging-clinical integration: PET/CT radiomics and clinical variables with KNN	Macro-average AUC: 0.960 in independent validation	Confirms the complementary role of PET metabolic features, CT anatomical features, and clinical variables	Requires standardized PET/CT acquisition and feature extraction protocols	n=896; SCLC =174 (19.4%); mixed lung cancer cohort; retrospective; center setting NR; training/test cohorts plus independent scanner-based validation (n=77); patient-level	Mixed-cohort evidence; 174 SCLC cases included
Overall survival prediction	Zheng et al.	CT images; clinical variables (non-imaging)	DL with imaging-clinical integration: CT survival signature plus clinical variables	C-index: 0.872–0.892 in multicenter cohorts	Provides relatively high-level evidence for CT-based survival prediction in SCLC	Retrospective multicenter design; prospective clinical utility remains unproven	n=1,061; SCLC =1,061 (100%); SCLC-only cohort; retrospective, four-center; training =543, internal validation =232, and two external validation cohorts (n=155 and 131); patient-level	Direct SCLC-specific evidence
OS prediction and PCI-associated survival outcomes	Zheng et al.	CT images and clinical features	DL/multimodal: CT survival signature plus clinical variables	C-index: 0.733–0.745; improved to 0.770–0.771 after adding clinical data	May help identify patients who benefit from prophylactic cranial irradiation and avoid overtreatment in low-risk patients	PCI decision-making requires prospective validation before routine clinical adoption	n=556; SCLC =556 (100%); SCLC-only cohort; retrospective, three-center; training =309, internal validation =133, and external validation =114; patient-level	Direct SCLC-specific evidence
OS and disease-free survival prognostication	Zhang et al.	H&E-stained pathological slides	DL pathomics: histopathology image model and PathoSig	HR for overall survival: 5.03; HR for disease-free survival: 2.63	Shows that pathomic features provide prognostic information beyond TNM staging and molecular subtyping	Requires high-quality pathological slides and standardized digital pathology pipelines	n=380; SCLC =380 (100%; pure or combined SCLC); retrospective, two-center; training =157, testing =39, two internal validation cohorts (n=44 and 46), and external validation =94; patient/slide-level	Direct SCLC-specific evidence
Pretreatment chemotherapy-response prediction	Vaidya et al.	Radiological and digital pathology images (cross-domain image integration)	Conventional radiomics/ML with radiology-pathology image integration	AUC: 0.78; superior to radiomics-only and pathomics-only models	Demonstrates that macroscopic imaging and microscopic pathological features provide complementary information for chemotherapy response prediction	Small cohort; retrospective design; not yet validated in treatment-guided trials	SCLC cohort n=50 (100%); SCLC-only subgroup; retrospective, single-center; patient-level Monte Carlo cross-validation repeated 100 times; no external validation	Direct SCLC-specific evidence based on a separate SCLC cohort
Pretreatment chemoimmunotherapy-response prediction	Nie et al.	CT/PET and pathology images; clinical and multi-omics data (non-imaging)	Multisource ML: radiomic, pathomic, clinical, and multi-omics features with random forest	AUC: 0.901	Suggests potential for predicting response to immunochemotherapy in extensive-stage SCLC	Complex multimodal data requirements may limit generalizability and clinical deployment	n=309; ES-SCLC =309 (100%); SCLC-only cohort; retrospective, single-center; internal training/validation split; patient-level; no external validation	Direct SCLC-specific evidence
9-month survival prognostication and molecular correlates	Jiang et al.	H&E-stained pathology images; genomic data (non-imaging)	ML pathomics: random survival forest signature	AUC: 0.85 for 9-month survival prediction	Links pathomic risk scores with CD8+ T-cell infiltration, MCL-1 amplification, and EP300 mutation	Needs biological validation and larger independent cohorts	n=118; ES-SCLC =118 (100%); SCLC-only cohort; retrospective, single-center; ten-fold feature-selection cross-validation and 1,000-resample bootstrap internal validation; patient-level; no external validation	Direct SCLC-specific evidence
SCLC lesion segmentation	Chen et al.	CT images	DL segmentation: hybrid 2D/3D CNN	Dice coefficient: 0.888; sensitivity: 0.872; precision: 0.909	Provides automated lesion delineation for subsequent radiomics and treatment planning	Segmentation robustness across scanners and institutions remains uncertain	n=134 CT scans; SCLC =134 (100%); SCLC-only cohort; single-center; random split into training =84, validation =20, and testing =30; scan-level; no external validation	Direct SCLC-specific evidence
Multicenter SCLC segmentation	Zheng et al.	Arterial-phase and venous-phase CT	DL segmentation: nnU-Net	Dice coefficient: 0.878 and 0.872	Establishes a benchmark for automated SCLC segmentation in multicenter cohorts	Requires evaluation in prospective and heterogeneous imaging settings	n=1,061; SCLC =1,061 (100%); SCLC-only cohort; retrospective, four-center; evaluation using arterial- and venous-phase test sets; scan-level; no prospective validation	Direct SCLC-specific evidence
PET/CT lung cancer segmentation	Xiang et al.	PET/CT images	DL segmentation: multimodal fusion and multiscale attention network	Dice coefficient: 0.803	Potentially transferable PET/CT segmentation method; SCLC-specific performance was not reported	Not limited to SCLC; SCLC-specific applicability requires further validation	Total n NR; SCLC n/proportion NR; cohort composition, study design, center setting, and validation type NR; image-level segmentation analysis	Indirect/transferable evidence; SCLC-specific sample size and performance NR

Treatment-related models predominantly used pretreatment baseline inputs; none evaluated serial on-treatment imaging for early response monitoring. NR, not reported; internal validation refers to random splitting, cross-validation, or bootstrap evaluation within the source dataset, whereas external validation refers to evaluation in an independent cohort. 2D/3D, two-dimensional/three-dimensional; AI, artificial intelligence; AUC, area under the curve; C-index, concordance index; CNN, convolutional neural network; CT, computed tomography; DL, deep learning; ES-SCLC, extensive-stage small cell lung cancer; FPN-SE-ResNet18, feature pyramid network with squeeze-and-excitation ResNet18; H&E, hematoxylin and eosin; HR, hazard ratio; KNN, k-nearest neighbors; LASSO, least absolute shrinkage and selection operator; MCL-1, myeloid cell leukemia 1; ML, machine learning; MRI, magnetic resonance imaging; NSCLC, non-small cell lung cancer; OS, overall survival; PCI, prophylactic cranial irradiation; PET, positron emission tomography; SCLC, small cell lung cancer; TNM, tumor-node-metastasis.

Analytical approaches were classified according to the final predictive model. Neural-network-based models were categorized as DL; pipelines using predefined or handcrafted features with regression, k-nearest-neighbor, support vector machine, or random forest algorithms were categorized as conventional radiomics/machine learning. Multimodal imaging was reserved for models combining two or more radiological imaging modalities. Radiology-pathology combinations were classified as cross-domain image integration, whereas models combining imaging with clinical variables, serum biomarkers, genomic profiles, or liquid-biopsy measurements were classified as imaging-non-imaging data integration. These integration strategies were considered independent of the model family. Automated segmentation alone was not considered sufficient to classify a downstream radiomics model as DL.

DL architectures and multimodal fusion strategies

Evolution and application of DL architectures

The architecture studies discussed in this section were conducted in heterogeneous lung-cancer populations. Their relevance to SCLC is classified in Table 2, while SCLC-specific clinical conclusions are based only on direct or mixed-cohort evidence with identifiable SCLC results. ResNet series models are among the most widely used architectures. ResNet18, ResNet50, and their variants are frequently employed as backbone networks or feature extractors in classification tasks (12-17). These architectures address optimization problems associated with deep networks, but their reported clinical performance cannot be generalized across heterogeneous datasets. Sui et al. evaluated ResNet18 in 192 patients with brain metastases, including 73 with SCLC and 119 with non-small-cell lung cancer (NSCLC). Patient-level five-fold cross-validation yielded a mean accuracy of 75.15%±4.0% and area under the curves (AUCs) of approximately 0.80, indicating moderate discrimination without independent external validation (12).

U-Net and its variants dominate the field of SCLC lesion segmentation. In a retrospective multicenter SCLC study, nnU-Net yielded Dice coefficients of 0.878 and 0.872 on arterial- and venous-phase CT test sets, respectively (18). In a separate dataset of 134 SCLC CT scans, the HSN network reported a mean Dice coefficient of 0.888 (19). CIs were not reported for these Dice coefficients, and the values should not be directly compared because they were obtained from different datasets and validation settings.

Attention mechanisms and hybrid architectures represent recent technological frontiers. Jacob et al. developed an attention-based CNN ensemble model and reported 98.8% accuracy using ten-fold cross-validation on publicly available PET/CT and CT datasets (20). However, the analysis was performed at the image-slice level, the patient-level sample size was not reported, and independent patient-level external validation was not performed; therefore, this value should not be interpreted as patient-level clinical accuracy. Hybrid CNN-Transformer models such as ConvFormer (21) and MS-GPTNet (22) exhibit superior performance in classification tasks by combining the local feature extraction capability of CNNs with the global context modeling ability of Transformers.

Generative adversarial network (GAN)-related methods have opened new avenues for cross-modal information fusion. Ma et al. proposed a conditional GAN (CGAN) that generates synthetic pathological features from CT images, effectively compensating for the lack of pathological details in non-invasive imaging (23). This approach enables pathological-level information acquisition without tissue biopsy.

The application of lightweight models such as MobileNetV2-CapsNet and MobileNet V3-Small reflects researchers’ focus on clinical deployment feasibility.

Multimodal fusion strategies

The shift from single-modal analysis to multimodal integration represents a key direction in SCLC research. Multimodal integration describes data fusion and does not by itself imply the use of DL. Existing studies involve imaging modalities including CT, PET/CT, MRI, and pathological images, while an increasing number of studies integrate clinical data, serum biomarkers, and genomic information into the analytical framework.

As a fundamental imaging modality, CT is used alone or as the core component of multimodal imaging combinations in approximately 25 studies. Its advantages include high prevalence, good standardization, high spatial resolution, and detailed anatomical structural information. CT is widely used either alone or as a component of multimodal models because of its availability, standardization, and spatial resolution. PET/CT integrates metabolic and anatomical information through strategies such as early fusion (24), late fusion (25), and cross-modal attention (21). Most PET/CT fusion studies cited here were conducted in NSCLC, general lung-cancer, or pulmonary-nodule datasets without separately reported SCLC outcomes (24-26). They therefore provide potentially transferable methodological evidence rather than direct validation of PET/CT fusion in SCLC.

Cross-domain integration of CT and digital pathology images represents an important form of image-based data integration. Jin et al. developed a contrastive constraint module to learn cross-modal associations between CT and pathological images during training, enabling near-pathological-level classification using only CT images at testing (27). Vaidya et al. constructed a conventional machine-learning model integrating radiological and pathological image features that significantly outperformed the corresponding single-source models in predicting chemotherapy response (28). The study by Jin et al. provides indirect lung-cancer evidence, whereas the chemotherapy-response analysis by Vaidya et al. was conducted in a separate cohort of 50 patients with SCLC and therefore provides direct SCLC-specific evidence.

Imaging-non-imaging data integration combines radiological or pathological image features with clinical variables, serum biomarkers, genomic profiles, or liquid-biopsy data. Clinical-imaging integration has improved model performance in several studies (29-31), whereas radiogenomics links imaging phenotypes with genomic characteristics. In 154 patients with limited-stage SCLC, Li et al. integrated clinical, radiomic, and genomic features and identified cyclin-dependent kinase 4 (CDK4) amplification and tumor mutational burden (TMB) status as predictors of chemoradiotherapy response (32), providing direct SCLC-specific evidence.

Analysis of clinical application effects

Diagnostic and classification efficacy

Precise diagnosis and subtype classification of SCLC constitute the most intensively researched area. In terms of task difficulty, common paradigms include three-class classification among SCLC, lung adenocarcinoma, and lung squamous cell carcinoma, while six-class or multi-class tasks pose significantly greater challenges.

Very high classification metrics were reported mainly in curated, image-level datasets. Jacob et al. reported 98.8% accuracy using ten-fold cross-validation of public PET/CT and CT image slices (20); however, the patient-level sample size and patient-disjoint data partitioning were not reported. If correlated slices from the same patient are distributed across folds, image-level cross-validation may overestimate generalizable performance. Tian et al. reported a mean five-fold-cross-validated accuracy of 98.84% on 750 public Kaggle images comprising benign lung tissue, lung adenocarcinoma, and lung squamous cell carcinoma, but the dataset contained no SCLC cases (33). Neither result, therefore, provides independently validated patient-level evidence for SCLC diagnosis, and CIs were not reported. However, model performance declines markedly in more clinically challenging tasks. Sui et al. achieved 75.2% accuracy in differentiating SCLC from NSCLC in brain metastases using MRI T1 images (12), Aydın et al. reported a three-class CT model with a sensitivity of 0.90 but a specificity of only 0.44 (34), highlighting the morphological overlap between SCLC and some NSCLC subtypes on CT. Chen et al. achieved 71.8% accuracy in a six-class cytological image task (15), indicating that increased complexity from expanded category numbers poses significant challenges to models.

Multimodal fusion demonstrates consistent value in diagnostic tasks. Wang et al. constructed a combined clinical-imaging-radiomic model that yielded a validation AUC of 0.967 (95% CI: 0.925–1.000), compared with 0.896 (95% CI: 0.826–0.967) for the clinical model and 0.836 (95% CI: 0.742–0.929) for the radiomics-only model (35). After integrating clinical features into a PET/CT radiomic model, Peng et al. achieved a macro-average AUC of 0.960 in an independent validation set, surpassing any single modality (36). These evidences indicate that information from different modalities is complementary, and multimodal fusion is an effective approach to improving classification efficacy.

Prognosis and survival prediction

Prognosis prediction is one of the areas supported by direct SCLC-specific evidence, including retrospective multicenter CT cohorts of 1,061 and 556 patients (18,37). Zheng et al. developed a deep-learning radiomics nomogram that yielded C-indices of 0.892, 0.873, and 0.872 in the internal validation cohort and two external validation cohorts, respectively (18). The model significantly outperformed the clinical nomogram (all P<0.05), whereas adding clinical factors did not significantly improve its performance (all P>0.05), suggesting that CT-derived features may capture key prognosis-related information. In another study by the same team, a DL overall survival prediction model achieved a C-index of 0.733–0.745 in 556 patients across three centers, which increased to 0.770–0.771 after combining clinical information (37).

Prognosis prediction based on pathological images also shows promising potential. Zhang et al. extracted 50 histomorphological phenotypic clusters from hematoxylin and eosin (H&E)-stained slides using comparative clustering, and in selected validation cohorts, high-risk PathoSig was associated with poorer overall survival (OS) [hazard ratio (HR) 5.03; 95% CI: 1.33–19.08; P=0.02] and DFS (HR 2.63; 95% CI: 1.46–4.71; P=0.001), providing independent prognostic information beyond tumor-node-metastasis (TNM) staging and molecular subtyping (13). Jiang et al. reported a 9-month time-dependent AUC of 0.845 for the random survival forest model, compared with 0.776 for the Cox model (P=0.018); the bootstrap mean AUC was 0.833 (0.761–0.905) (38).

Notably, prognostic predictive efficacy varies across studies. Li et al. constructed a radiogenomic model based on LASSO feature selection and Cox regression that achieved a C-index of 0.574–0.656 in predicting progression-free survival in patients with limited-stage SCLC receiving definitive chemoradiotherapy (32). The relatively low predictive efficacy may be related to treatment specificity, small sample size (154 cases), and different definitions of endpoint events. This suggests that the efficacy of prognosis prediction models should be interpreted in the context of specific clinical scenarios.

Treatment response and treatment-associated outcomes

Studies in this section addressed distinct endpoints and were therefore interpreted separately, including response to chemotherapy, chemoradiotherapy, or chemoimmunotherapy; survival outcomes such as PFS and OS; PCI-associated survival outcomes; and exploratory molecular correlates (28,32,37-39). Most models used pretreatment baseline data and therefore represented pretreatment prediction or prognostication rather than early response monitoring.

In chemotherapy response prediction, Vaidya et al. constructed a conventional machine-learning model integrating handcrafted radiomic and pathomic features that predicted SCLC chemotherapy response with an AUC of 0.78, significantly superior to the radiomics-only model (0.74, P<0.034) and pathomics-only model (0.64, P<0.012) (28). This finding indicates that macroscopic imaging features and microscopic pathological features provide complementary information in reflecting treatment sensitivity.

Benefit stratification for prophylactic cranial irradiation (PCI) is a key issue in SCLC treatment decision-making. Zheng et al. used a DL overall survival signature to stratify patients into high- and low-risk groups. PCI was associated with improved survival in the high-risk groups but not in the low-risk groups (37). However, this retrospective association does not establish treatment-benefit prediction, and prospective validation is required before the model can inform PCI decisions.

In immunochemotherapy response prediction, Nie et al. constructed a multisource random forest model integrating radiomic, pathomic, clinical, and multi-omics features to predict chemoimmunotherapy response in patients with extensive-stage SCLC patients with an AUC of 0.901, identifying independent predictors including energy, sum of squares (SOS), mean sum (MES), sum variance (SUV), sum entropy (SUE), difference variance (DIV), and pathomic scores (39). Jiang et al. linked pathomic risk scores to CD8+ T-cell infiltration, MCL-1 amplification, and EP300 mutation (38), providing clues for understanding the association between pathological morphological features and the tumor immune microenvironment.

Segmentation performance

Accurate lesion segmentation is the foundation for subsequent radiomic analysis. For SCLC lesion segmentation, nnU-Net achieved Dice coefficients of 0.878 and 0.872 on arterial- and venous-phase CT test sets, respectively (18). Chen et al. reported a Dice coefficient of 0.888, sensitivity of 0.872, and precision of 0.909 using the HSN network in 134 patients with SCLC (19). CIs were not reported for these metrics. Because lesion size, annotation procedures, scanner characteristics, and validation design differed between studies, the small numerical differences in Dice coefficients should not be interpreted as clinically meaningful. Both studies nevertheless provide direct SCLC-specific segmentation evidence. By contrast, the PET/CT segmentation study by Xiang et al. did not report the number of SCLC cases or SCLC-specific performance and was therefore considered indirect, potentially transferable evidence (40). A summary of the representative studies and their relevance to SCLC is provided in Table 2.

Thematic analysis and synthesis

Evolutionary trends in imaging and data integration strategies

Across the reviewed studies, imaging and data-integration strategies have evolved from late fusion of separately modeled data sources to feature-level and deep fusion. Early studies mostly adopted late fusion, i.e., fusing decision outcomes after training individual modal models; this method is simple and direct but fails to fully exploit cross-modal interactive information. Feature-level fusion retains more information by concatenating or weighting intermediate feature representations of different modalities but also increases the risk of overfitting.

For multimodal imaging, attention-based cross-modal fusion has become an important strategy for integrating complementary information from different imaging modalities. Such methods dynamically focus on cross-modal correlated regions with the richest information by calculating attention weights between different modal features. Tao et al. constructed a 3D ConvFormer with a dual-branch cross-modal feature interaction module, achieving 89.25% accuracy in 3,173 patients (21). Xiang et al. designed MFMA-SegNet, which achieved a Dice coefficient of 0.803 in PET/CT segmentation tasks (40). Because these studies were conducted in broader lung-cancer populations without separately reported SCLC outcomes, they are discussed as potentially transferable fusion strategies rather than as SCLC-specific clinical evidence.

More innovative is the cross-modal information generation strategy. Ma et al. generated synthetic pathological features from CT images via conditional GAN (23), Jin et al. established associations between CT and pathological images during training using a contrastive constraint module, enabling enhanced classification using only CT images at testing (27). These methods attempt to fundamentally solve the problem of pathological access gaps and hold important clinical translation value.

SCLC-specific biological associations

AI-based imaging and pathomic analyses can provide predictive models and offer insights into SCLC biology. By linking quantitative imaging or pathological features with molecular data, researchers have identified several SCLC-specific biological associations. These statistical associations are hypothesis-generating and do not establish biological mechanisms without independent or functional validation. Jiang et al. found that high pathomic risk scores were associated with more abundant stromal CD8⁺ T-cell infiltration, higher probabilities of myeloid cell leukemia 1 (MCL-1) amplification, and EP300 mutation (38). These associations suggest potential relationships between histomorphological features, the tumor immune microenvironment, and genomic alterations but do not establish causal mechanisms or treatment benefit.

Li et al. identified CDK4 amplification, GATA6, MAPK/ERK pathway genes, and TMB status as predictors of response to definitive chemoradiotherapy in limited-stage SCLC through radiogenomic analysis (32). These findings connect imaging phenotypes to molecular characteristics, providing candidate biomarkers for treatment stratification. However, direct SCLC radiogenomic evidence remains limited, with the strongest recent evidence derived from retrospective integration of pretreatment clinical, radiomic, and genomic data; external validation and functional confirmation remain necessary.

In a mixed cohort containing 24 SCLC cases among 345 patients, SHapley Additive exPlanations (SHAP) analysis indicated that neuron-specific enolase (NSE) contributed to SCLC classification, whereas Lung Imaging Reporting and Data System (Lung-RADS) contributed to adenocarcinoma classification (41), these findings require external validation. Such analyses help describe model behavior but do not validate biological mechanisms.

The biological interpretability of imaging-based DL models should also be discussed in the context of the emerging molecular taxonomy of SCLC. SCLC is no longer regarded as a homogeneous neuroendocrine malignancy; instead, transcription factor-defined subtypes characterized by dominant achaete-scute family bHLH transcription factor 1 (ASCL1), neuronal differentiation 1 (NEUROD1), POU class 2 homeobox 3 (POU2F3), or Yes-associated protein 1 (YAP1) programs have been proposed, and subsequent studies further identified an inflamed subtype with enhanced immune pathway activation (42-44). This classification framework is clinically relevant because different subtypes may exhibit distinct therapeutic vulnerabilities, including differential dependence on neuroendocrine lineage programs, DNA damage response pathways, immune infiltration, and antigen-presentation status (42-46). For example, the inflamed subtype appears to be associated with stronger immune signaling and may be more likely to benefit from chemoimmunotherapy, whereas other subtypes may require lineage-specific or targeted therapeutic approaches (43,44). These findings provide a biological rationale for integrating imaging features with pathological, transcriptomic, and genomic data in SCLC.

In this context, radiogenomics may become a key strategy for converting “black-box” imaging models into biologically meaningful decision-support systems. If CT, PET/CT, or MRI phenotypes can be linked to transcriptional subtypes, tumor mutational burden, immune microenvironment features, DLL3 expression, or DNA damage repair alterations, imaging-derived features could be evaluated as candidate biomarkers for stratification in future treatment studies (43,46,47). Tarlatamab, a DLL3-targeted bispecific T-cell engager, has shown durable antitumor activity in previously treated SCLC, suggesting that future imaging models may need to predict not only conventional chemotherapy response but also eligibility for emerging mechanism-based therapies (47). However, such applications require careful validation because imaging phenotypes may reflect mixed signals from tumor cellularity, necrosis, stromal components, vascularity, and treatment-induced inflammation. Therefore, future studies should combine explainable AI, spatial pathology, and multi-region molecular sampling to clarify whether model-derived features represent true tumor biology or merely technical correlations.

Challenges and pathways for clinical translation

Despite encouraging technical progress, the clinical translation of AI and radiomics models using imaging or multisource data faces multiple obstacles. SCLC-specific datasets are generally small, and SCLC is frequently a minority class in mixed lung-cancer datasets. Small-sample training limits model generalization and increases the risk of performance degradation in multicenter validation.

In terms of study design, retrospective single-center studies predominate (35,38,39), lacking prospective validation (13,28). The nature of models remains a major barrier to clinician acceptance. Although techniques such as SHAP analysis (48,49), Grad-CAM visualization (50), and interpretable auxiliary branches (51) attempt to improve transparency, gaps still exist in meeting clinical requirements for interpretability.

Reported performance values should not be compared directly across studies without considering cohort composition and validation design. Very high accuracy or AUC values may be inflated in small, selected, or class-imbalanced cohorts. Image-level splitting can also lead to non-independent evaluation when correlated slices or tiles from the same patient occur in both training and validation folds. Internal random splitting and cross-validation estimate performance within the same data source but do not demonstrate transportability to new institutions, scanners, or patient populations. Public or highly curated datasets may contain cleaner images and greater interclass separation than routine clinical data. Performance should therefore be accompanied by patient and class counts, the unit of analysis, patient-level data partitioning, external validation, class-specific sensitivity and specificity, calibration, and 95% CIs where available. The TRIPOD, TRIPOD+AI, and CLAIM guidelines provide appropriate frameworks for such reporting (52-54). The TRIPOD statement and its AI-focused update emphasize transparent reporting of prediction model development and validation, including clear descriptions of candidate predictors, outcome definitions, sample size considerations, model specification, and performance measures (52,53). For medical imaging AI, the CLAIM checklist further recommends detailed reporting of imaging acquisition, preprocessing, annotation, data partitioning, ground truth definition, model architecture, training procedures, and failure analysis (54,55). Incorporating these frameworks would substantially improve the credibility and reproducibility of future SCLC imaging studies.

Radiomics-specific standardization is equally important. The Image Biomarker Standardisation Initiative demonstrated that radiomic feature definitions and software implementations can vary, and standardized feature calculation is necessary to improve reproducibility across institutions (56). In SCLC, this issue may be particularly important because tumors often shrink rapidly after treatment, and small segmentation differences can markedly influence texture features, metabolic tumor volume, or longitudinal delta-radiomic parameters. Therefore, future studies should report scanner type, reconstruction kernel, voxel spacing, contrast phase, PET uptake time, segmentation protocol, feature extraction software, harmonization method, and test-retest stability. In addition, multicenter studies should avoid random patient-level splitting when multiple scans from the same patient are included, because this may cause data leakage and overestimate performance. Model evaluation should include external validation across hospitals, scanners, and treatment eras; calibration plots and clinical utility analyses should be reported alongside discrimination metrics; and subgroup performance should be assessed according to stage, treatment regimen, metastatic site, and image quality. These methodological safeguards are necessary before any model can be considered for prospective clinical testing. No SCLC-specific AI model has yet been reported as receiving regulatory approval or being integrated into a clinical trial, indicating that clinical translation remains at an early stage.

Future prospects

Figure 1 summarizes the current evidence landscape and the major priorities for future SCLC research, including prospective evaluation, standardized model development, clinical integration, and imaging-non-imaging data integration.

Figure 1.

Current evidence landscape, reported biological associations, and proposed research directions for AI and radiomics in SCLC. Red numerical values represent performance reported in individual studies and should be interpreted according to cohort size and validation design. Arrows in the biological-association panel indicate reported statistical associations rather than established biological mechanisms. The figure was created by the authors using bioGDP (biogdp.com). AI, artificial intelligence; AUC, area under the curve; CD8, cluster of differentiation 8; CDK4, cyclin-dependent kinase 4; CNN, convolutional neural network; CT, computed tomography; MCL-1, myeloid cell leukemia 1; MRI, magnetic resonance imaging; PCI, prophylactic cranial irradiation; PET/CT, positron emission tomography/computed tomography; SCLC, small cell lung cancer; TMB, tumor mutational burden.

A clinically oriented roadmap for SCLC AI and data-integration research should proceed in three stages, moving from retrospective model consolidation to prospective observational validation and finally to interventional AI-guided trials (Figure 2). The first stage is retrospective model consolidation, in which existing CT, PET/CT, MRI, pathology, and clinical datasets are harmonized, re-annotated, and evaluated using unified reporting standards. In this stage, researchers should prioritize clinically meaningful endpoints such as early progression after first-line chemoimmunotherapy, brain metastasis risk, PCI benefit, thoracic radiotherapy response, and survival after relapse. The second stage is prospective observational validation. AI outputs should be generated before treatment decisions but not used to guide management initially, allowing investigators to determine whether model predictions remain accurate in contemporary treatment settings and whether they add value beyond standard clinical variables. DECIDE-AI provides a useful framework for early-stage clinical evaluation of AI-based decision-support systems, especially regarding human-AI interaction, workflow integration, error analysis, and safety monitoring (57). The third stage is interventional evaluation, in which AI-assisted stratification is tested in prospective trials. Prospective studies should evaluate whether model-defined risk groups can improve brain surveillance, PCI-related decisions, or clinical-trial selection.

Figure 2.

Proposed clinical translation roadmap for AI, radiomics, and imaging/data integration in SCLC. The three stages represent proposed research priorities rather than established clinical pathways. The figure was created by the authors using bioGDP (biogdp.com). AI, artificial intelligence; QA, Quality Assurance; SCLC, small cell lung cancer.

Future evaluation should also address fairness across clinically relevant subgroups and feasibility in resource-limited settings. AI should be positioned as a complementary decision-support tool, with calibration, safety, and human oversight evaluated before clinical use.

Conclusions

The evidence supporting AI and radiomics based on multimodal imaging or imaging-non-imaging data integration in SCLC is heterogeneous. Direct SCLC-specific evidence is currently strongest for prognosis prediction, treatment-response prediction and treatment-associated outcomes, and CT-based segmentation, whereas diagnostic classification is supported mainly by mixed-cohort studies with identifiable SCLC results. Studies based on broader lung-cancer populations without separately reported SCLC outcomes provide only indirect, potentially transferable evidence. Therefore, performance improvements from integrating multiple imaging or non-imaging data sources should be interpreted within the specific cohorts and clinical tasks in which they were demonstrated. Imaging-based deep-learning and conventional radiomics/machine-learning models for treatment-response and PCI-associated outcomes remain investigational because prospective model-guided evidence is lacking. Overall, prospective multicenter validation and transparent reporting remain necessary before clinical implementation.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.