Skip to main content
NCBI home page
Chemical & Biomedical Imaging logoLink to Chemical & Biomedical Imaging
. 2026 Apr 3;4(9):2091–2101. doi: 10.1021/cbmi.6c00025

Lipid Metabolism-Related Rapid Pathological Diagnosis and Staging of Breast Tissue Sections Using Two-Photon Fluorescence Imaging Assisted with a Carbonized Polymer Dot Probe

Zitong Yu †, Jiaqi Wang †, Ying Liu ‡, Weiqing Xu †, Bing Han ‡,*, Shuping Xu †,§,∥,*
PMCID: PMC13625745  PMID: 42820010

Abstract

Rapid and accurate intraoperative pathologic diagnosis is essential for intraoperative decision-making, improving patient prognosis, and avoiding reoperation. This study developed a unique rapid pathological diagnosis and staging technique for intraoperative tissue sections based on lipidomic relevance, using two-photon fluorescence (TPF) imaging assisted by a lipid-targeted carbonized polymer dot (CPD) probe (∼8 nm) and multiple classification algorithms, including machine learning and deep convolutional networks. This platform enables ultrabright TPF imaging in response to lipid microenvironments and can be completed within 10 min. Its feasibility was validated through rapid diagnostics of intraoperative frozen sections of breast tissues. Due to abnormal levels of lipid metabolism, normal breast tissues and pathologically staged I–III cancerous tissues were well distinguishable, and the sensitivities and specificities for pathologic diagnosis were all above 90%. The combination of CPD and TPF microscopy provides a rapid and accurate method for detecting lipidomes in pathological frozen sections, marking a significant advancement over hematoxylin and eosin (H&E)-staining-based pathological diagnosis in terms of speed, ease of operation, and relevance to chemical components. This study is vital for advancing lipidome-related research and enhancing the understanding of the lipidomic atlas.

Keywords: Carbonized polymer dot, intraoperative rapid pathology, two-photon fluorescence, machine learning, convolutional neural networks

graphic file with name im6c00025_0010.webp

graphic file with name im6c00025_0008.webp

1. Introduction

Identifying cancer cells, assessing their differentiation, and providing decisive advice to clinicians are daily challenges for pathologists. Currently, the gold standard for the pathological diagnosis of biopsy and operative samples primarily relies on traditional (hematoxylin and eosin, H&E) staining, followed by microscopic examination by a skilled pathologist. , The morphological features of cell nuclei highlighted by H&E staining serve as the diagnostic basis. The H&E staining process is intricate, involving multiple detailed steps, including tissue fixation, dehydration, and embedding. Each step requires precise control and timing to maintain staining quality and effectiveness. − For instance, the duration of hematoxylin staining must be adjusted based on laboratory temperature, and the counterstaining duration depends on the visual outcomes in cell nuclei. Moreover, such diagnostic methods primarily rely on the pathologist’s subjective judgment, require high expertise, and often result in low diagnostic efficiency and significant human-factor bias. Given these challenges, there is a significant need for computer-assisted histopathology that not only alleviates the pathologists’ workload but also enhances diagnostic objectivity and efficiency. Especially for intraoperative diagnosis, a rapid and accurate pathological diagnostic method is pivotal, which will facilitate immediate surgical decisions, improve patient outcomes, and reduce the likelihood of reoperation.

Aside from morphological differences, one viable strategy to distinguish tumors from normal tissues is based on differences in chemical composition, such as lipidomes arising from disordered lipid metabolism. Lipid metabolism plays an essential role in cellular energy regulation, particularly in cancer cells, which alter their metabolic pathways to support the synthesis of adenosine triphosphate (ATP) and vital macromolecules necessary for their proliferation, division, and survival. − Lipids, a diverse group of biomolecules, are foundational to the structure of biological membranes and function as both signaling entities and energy sources. − Within the organism, these lipids undergo synthesis, catabolism, and other metabolic transformations. Studies have shown that cancer cells manipulate lipid metabolism to foster their rapid growth, survival, and invasive capabilities, including migration and metastasis. An increase in lipid synthesis or absorption significantly accelerates tumor growth and development. , Cancer cells can not only synthesize lipids themselves but also absorb exogenous lipids to meet their demand for unsaturated fatty acids. Aberrant lipid metabolic signaling has the potential to alter the behavior of both malignant and nonmalignant cells in tumor microenvironments, which may lead to the onset of cancer, its progression, spread, recurrence, and diminished response to treatment. Existing observations suggest that abnormalities in lipid metabolism and excessive lipid accumulation are crucial for cancer cell survival. , Therefore, variations in lipid metabolism levels across different tumor stages can serve as a marker for classifying cancer progression.

To effectively extract information about lipidomes in tissue samples for a new, practical pathological diagnostic technique, this study proposes a fast, impersonal detection platform for rapid intraoperative pathological diagnosis of tissue sections using two-photon fluorescence (TPF) imaging aided with a lipid-specific, carbonized polymer dot (CPD)-based nanoprobe and multiple image classification algorithms. TPF imaging offers significant advantages over single-photon fluorescence imaging, including deeper tissue penetration, higher spatial resolution, reduced background fluorescence, and less tissue damage, − making it ideal for studying highly integrated deep tissues and living systems. − However, there are not many optional TPF probes. Our study employed a CPD-based nanoprobe that enables wash-free staining and high-performance TPF in response to lipid environments. , Using this lipid-specific CPD probe, we examined human breast cancer (BC) pathological sections at different stages. They were quickly stained with the CPD probe (5 min), followed by examination with TPF microscopy. Machine learning and convolutional neural networks (CNNs) were utilized to process the TPF images to assist in staging BCs. This platform serves as a reliable and applicable method for intraoperative rapid pathology and tumor diagnosis, leveraging lipidomic metabolism distinct from traditional nuclear morphology-based pathology, and paves the way for pathological staging (Scheme ).

1. A Workflow for Rapid Intraoperative Pathological Diagnosis Related to Lipid Metabolism .

1

Open in a new tab

a The top plate shows the traditional intraoperative pathological diagnostic procedure. The bottom panel shows our rapid pathologic diagnostic process and intraoperative staging of pathological sections based on lipidomic metabolism.

2. Experimental Section

2.1. Synthesis of CPD

The CPD-based nanoprobe with TPF and lipid-responsive features was prepared by solvothermal synthesis using N,N-diethyl-p-phenylenediamine as a precursor and ethanol as the solvent in a Teflon-lined stainless steel vessel autoclave, followed by purification via column chromatography (Figure S1). − More details about the synthesis of CPD are provided in Supporting Information.

2.2. Lipid-Targeting Feature of CPD Explored by Cellular Co-Localization Assay

To evaluate the lipid-targeting ability of our CPD, we conducted cellular colocalization assays. MCF-7 cells were stained with two commercial dyes: BODIPY 493/503 for LD localization and Hoechst 33342 for nucleus localization. The cells were cocultured with these dyes at 37 °C for 30 min. To remove unbound dyes, we washed the cells three times with phosphate-buffered saline (PBS, pH 7.4). Cells were then stained with the CPD probe (50 μg/mL) for 10 min and subsequently imaged without further washing steps.

Image acquisition was performed using a Nikon A1 confocal laser scanning microscope (CLSM, Nikon Co. Ltd., Japan). CPD: λex = 633 nm; BODIPY 493/503: λex = 488 nm; Hoechst 33342: λex = 405 nm. The Pearson correlation coefficients between BODIPY 493/503 and the different CPD colors were calculated directly using ImageJ software.

2.3. Experiments for Pathological Sections

The frozen tissues were obtained from the Department of Breast Surgery, the First Hospital of Jilin University (Ethics approval number: 2022–449). This study intends to collect tissue samples from BC patients. The patient’s informed consent is obtained and signed. Surgical resection of invasive BC was performed on twenty-three women, with patient age and pathological details summarized in Table S1. Fresh cancerous and corresponding normal tissue samples were procured from these patients, with the latter obtained from regions at least 5 cm away from the tumor margins.

BC tissues and adjacent normal tissues were homogenized using a tissue homogenizer (Bertin-Precellys 24) in RIPA lysis buffer containing 1× protease inhibitor cocktail and 1 mM PMSF. The lysates were centrifuged at 12,000 × g for 15 min at 4 °C, and the supernatants were collected. Protein concentrations were quantified using a BCA Protein Assay Kit. Total cellular RNA was extracted using the RNeasy Plus Mini Kit.

Other samples were rapidly processed by the pathology department for cryoembedding and sectioning using a frozen slicer (Leica-CM3050S). The initial 5 μm sections were stained with H&E for diagnostic confirmation and to ascertain cancer cell density. Subsequent 15 μm sections were prepared for CPD staining and TPF imaging. The institutional review committee approved the study protocol, and written informed consent was obtained from all participants.

The Tumor Node Metastasis (TNM) staging was verified using H&E staining. As displayed in Scheme , Paraffin-embedded tissue sections were initially deparaffinized in xylene for 5–10 min, followed by hematoxylin staining for 5 min. After washing with water, the sections were briefly immersed in 0.5% hydrochloric acid for 1–3 s, rinsed with water for 1 min, and subsequently blued in 0.2% ammonia–water for 5–10 min. They were stained with eosin Y solution for 10–15 s and dehydrated twice in 95% alcohol for 30 min each. The thickness of H&E-stained sections was 8 μm for normal tissue and 6 μm for cancerous tissue. Detailed staining procedures for H&E are available in the support information.

2.4. TPF Imaging

For TPF imaging of the pathological frozen sections, tissue slices were incubated in PBS containing 100 μg/mL of CPDs at 37 °C for 5 min. After three cycles of PBS washing, TPF imaging was performed using a multiphoton laser-scanning confocal microscope (Olympus, FVMPE-RS, Japan) with a high-speed scanning galvanometer. The setup for TPF imaging is displayed in Figure S5. The laser source was a tunable picosecond laser (Micro Emerald, APE, Germany, λex = 790 nm). The optical system included a 10× objective lens with a numerical aperture (NA) of 0.9, optimizing the excitation and collection of TPF signals in upright mode. All operational parameters were controlled using the FV31S-SW software.

For imaging of the pathology H&E sections, we switched the same setup to a bright-field imaging mode and changed the mode from “LSM” to “Ocular.” All operational parameters were controlled using the CellSens Standard software.

2.5. Image Data Analysis

RStudio software was utilized to process 417 fluorescence images in an R 4.2.2 environment. Statistics of average fluorescence intensity were generated using the EBImage package, which converted each image to grayscale, extracted pixel values, and calculated the average grayscale intensity and standard deviation. Pixel distribution maps were created using the hist function of the EBImage package. To standardize the process, pixel distributions across all images were uniformly spaced at intervals of 0, 0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.35, 0.40, 0.45, and 0.50, forming a pixel distribution matrix.

Machine learning algorithms were employed to differentiate the stages of pixel distribution. The Rtsne package was used to downscale the pixel distribution matrix with a perplexity of 15. A random forest classification model was constructed for four stages using the randomForest package, with a training-to-test set ratio of 8:2. The pheatmap package was used to generate a grouped clustering heatmap. Construction of the image data classification model involved using the Keras and TensorFlow packages to develop a CNN model for four-stage classification, directly processing image data with a training-to-test set ratio of 8:2, batch size of 10, and 30 epochs. Ten random repetitions were conducted using different training-and-test sets to validate the model’s repeatability. Model compilation involved the Adam optimizer and categorical_crossentropy as the loss function, with accuracy and cross-entropy loss, which can measure the model’s fit and precision and assess the discrepancy between the trained probability distribution and actual data, respectively.

3. Results and Discussion

3.1. Characterizations of CPD and Its Lipid-Targeting Feature

The particle size and morphology of the prepared CPD were examined using transmission electron microscopy (TEM). As shown in Figure A, the CPD dispersed in ethanol appeared as uniformly spherical nanoparticles. The average diameter of these particles (Figure B) was 7.87 ± 0.53 nm. X-ray photoelectron spectroscopy (XPS) was utilized to provide detailed information on the surface chemistry of the CPD. Figure C illustrates characteristic peaks at 284.93, 399.31, and 531.98 eV, corresponding to C 1s, N 1s, and O 1s, respectively. The high-resolution XPS spectra (Figure D–F) reveal that the C 1s peak could be deconvoluted into three components: sp 2-bonded carbon at 284.80 eV, carbon–nitrogen interactions at 285.78 eV, and a carbon group at 288.20 eV. The N 1s spectrum exhibits two peaks at 399.43 and 401.04 eV, corresponding to amino- and pyrrole-type nitrogen, respectively. The O 1s scan is resolved into peaks at 531.89 and 533.78 eV, associated with C–O single bonds and C–O–C or C–OH groups, respectively. Quantitative analyses of the elemental and functional group ratios are presented in Tables S2–S5. Figure G shows the emission spectra of CPD dissolved in ethanol. It can be observed that it is excitation wavelength-independent, giving an individual fluorescence emission band at 630 nm.

1.

1

Open in a new tab

A TEM image (A), a histogram of the particle size distribution (B), an XPS survey scan (C), and the scan ranges of C 1s (D), N 1s (E), and O 1s (F) of the CPD, respectively. (G) Emission spectra of the CPD at different excitation wavelengths. All spectra were recorded in ethanol. (H) The fluorescence spectra of the CPD dispersed in water and n-octanol with an excitation wavelength of 560 nm. The inset shows a photograph of CPD dispersed in water and n-octanol phases under a UV light (375 nm). (I) Fluorescence spectra of the CPD in PBS, O/W emulsion, DNA, RNA, BSA, HSA, ATP, and ADP solutions, respectively.

The significant difference between the hydrophobicity and hydrophilicity of a probe is the basis for designing a lipid-targeting probe. Lipophilicity, the tendency of compounds to partition preferentially in a nonpolar lipid medium and not in the aqueous phase, was assessed using a conventional two-phase partitioning method in mutually immiscible nonpolar and polar liquids. The lipid-responsive properties of CPD were first examined by dissolving it in n-octanol to mimic an intracellular lipid-like environment. As shown in Figure H, a significant change in fluorescence between two solvents was observed upon exposure to UV light, as supported by fluorescence spectroscopy. This suggests that our CPD exhibits strong fluorescence in lipid-like environments, preliminarily proving its lipid-targeting capabilities. The fluorescence feature of CPD is attributed to a particular D-π-A structure that is environmentally sensitive and can realize polarity sensing of microenvironments.

To investigate whether CPD interacts with lipid environments in vitro, an oil/water (O/W) emulsion of glycerol trioleate in a PBS buffer solution was used to simulate an LD-contained climate. Besides, interactions with DNA, RNA, bovine serum albumin (BSA), human serum albumin (HSA), ATP, and adenosine diphosphate (ADP) served as controls. The fluorescence spectra of CPD were analyzed after its interaction with various environments. The experimental results (Figure I) indicate that adding the O/W emulsion to a PBS solution significantly enhanced the fluorescence intensity of CPD. In contrast, adding other comparative references did not produce a notable change in fluorescence intensity. This suggests that CPD emits unique fluorescence once it meets lipid environments, and the improvement in fluorescence is attributed to its vesicle-like structure in O/W emulsions.

The LD imaging ability of CPD in living MCF-7 cells was verified. Before the CPD was applied to cells, its cell toxicity was assessed by the WST-1 assay. The results are shown in Figure S2, with a survival rate exceeding 85%, demonstrating good biocompatibility and minimal cytotoxicity of our CPDs. The LD specificity of CPDs was confirmed using the commercial organelle localization dye BODIPY 493/503 and confocal fluorescence microscopy, as depicted in Figure A. The correlation between fluorescence intensity and pixel position for CPDs in BODIPY 493/503 (Figure S3A–C) demonstrates almost identical curves, suggesting the high LD targeting ability of the CPD. A high Pearson’s colocalization coefficient (PCC) of 0.912 was achieved. Co-incubation imaging results indicate that these CPDs effectively target intracellular LDs. Additionally, the CPDs’ LD-targeting ability was further evaluated in LD-enriched MCF-7 cells, in which oleic acid was used to induce high LD expression (Figure B). Confocal laser-scanning microscopic images (CLSM) reveal a gradual increase in cell fluorescence as the concentration of oleic acid increases. This illustrates the significant sensitivity of our CPD to lipid content and its potential for tracking changes in abnormal lipid metabolism in cells.

2.

2

Open in a new tab

(A) CLSM images of MCF-7 cells costained with Hoechst 33342 (living nuclear dye), BODIPY 493/503, CPD, and their merged images. Scale bar: 20 μm. (B) CLSM images of MCF-7 cells pretreated without (control) or with different concentrations of oleic acid (OA) and then incubated with CPD for 10 min, respectively. Hoechst: λex = 405 nm; CPD: λex = 561 nm; BODIPY: λex = 488 nm. Scale bar: 20 μm.

3.2. Abnormalities of Lipid Metabolism in BC Tissues

Lipid metabolism is a fundamental part of metabolic reprogramming in tumors. Cancer cells disrupt normal lipid metabolism by increasing exogenous lipid uptake and de novo synthesis of fatty acids and cholesterol. Disruption in fatty acid metabolism is linked to the oncogenic transformation in various cancers, such as BC. − Studies also pointed out the use of exogenous lipids (e.g., oleic acid) by breast cancer cells to regulate the direction of lipid metabolism, as well as ab initio fatty acid synthesis. , In addition, many proteins are involved in the regulation of lipid metabolism in cancer cells, controlling cellular uptake of fatty acids involved in energy metabolism. For example, CD36 mRNA expression significantly reduces exogenous fatty acid uptake. In addition to this, stearoyl coenzyme A desaturase (SCD1), which converts saturated fatty acids to monounsaturated fatty acids, has been found in all cell types as well as in all plants and animals. Research indicates that an increase in SCD1 is observed in various types of tumor cells and contributes to cancer progression. − Therefore, to verify abnormal lipid metabolism in breast cancer, we explored the CD36 and SCD1 gene expressions in cancerous tissues versus adjacent normal tissues in breast cancer patients. CD36 and SCD1 levels were determined in BC and adjacent normal tissues by Western blot. Fresh postoperative tissues were obtained from the Department of Breast Surgery, the First Hospital of Jilin University (Ethics approval number: 2022–449). Cell lysates were immunoblotted with the indicated antibodies. β-Actin was used as an internal control. As shown in Figure A (the uncropped blots are provided in Figure S4), Western blot analysis reveals the upregulation of CD36 and SCD1 in BC tissues compared to paraneoplastic normal tissues, suggesting that they may be involved in lipid metabolic reprogramming and tumor progression. In addition, the levels of CD36 and SCD1 mRNAs in BC and paraneoplastic normal tissues were detected by real-time fluorescence quantitative polymerase chain reaction (qPCR) using 36B4 as an internal reference gene. As shown in Figure B, the qPCR results reveal a significant upregulation of SCD1 and CD36 mRNA levels in BC tissues compared to adjacent normal tissues, further corroborating their dysregulated expression at the transcriptional level. In summary, both Western blot and qPCR demonstrate that it is feasible to clinically diagnose BC by examining metabolic abnormalities in BC tissues. This strengthens the foundation of our design and may help identify new targets for metabolic therapy of BCs.

3.

3

Open in a new tab

(A) CD36 and SCD1 protein levels were determined in BC and adjacent normal tissues, analyzed by Western blot. Cell lysates were immunoblotted with the indicated antibodies. β-Actin was used as an internal control. The results of three replicated experiments. (B) CD36 and SCD1 mRNA levels were determined by real-time qPCR in BC and adjacent normal tissues. 36B4 was used as the internal reference gene. *: p < 0.05, **: p < 0.01, ***: p < 0.001. All statistical analyses comparing intergroup differences in this study employed a two-tailed independent samples t test for pairwise comparisons between two groups and a one-way analysis of variance for pairwise comparisons among multiple groups.

Oil Red O staining is a histopathological staining technique that uses Oil Red O dye to specifically visualize neutral lipids within tissues. It is primarily applied in pathological diagnosis and research fields. The principle behind this technique is based on the lipophilic nature of Oil Red O, which can bind to lipid components such as triglycerides to produce a color reaction. Similarly, we performed Oil Red O staining experiments on breast cancer tissue and adjacent normal tissue, with the results shown in Figure S6. The results indicate that in breast cancer tissue samples, there are a large number of red-stained lipid particles in the cytoplasm of ductal epithelial cells, necrotic areas, and the intracellular spaces of surrounding cancer cells. In contrast, in adjacent normal tissue, only a small number of red-stained lipid particles are present in the glandular and stromal components. This demonstrates that lipid content is higher in breast cancer tissue than in adjacent normal tissue. This validates the specificity of the probe for lipid staining in tissue sections.

3.3. Staging Tumor Grades by TPF Intensity

Owing to the excellent targeting performance of the probe for lipid imaging and the excellent penetration ability of TPF, we employed this CPD-based nanoprobe for rapid intraoperative pathological diagnosis of BCs. Frozen postoperative tissue sections obtained from the Department of Breast Surgery, the First Hospital of Jilin University (Ethics approval number: 2022–449) were used. Corresponding pathological staging results are listed in Table S1. It should be mentioned that the rapidly frozen sections of breast tissues were adopted because they retain the intact lipid components relative to paraffin sections. Two adjacent frozen sections were collected for each tissue. One was for H&E staining, and this section was initially imaged to help us determine cell locations and nuclear shapes. The other was stained with the CPD-based probe for TPF imaging. A total of 417 images of frozen sections of BC tissues were analyzed. H&E staining, which is widely considered the “gold standard” for grading pathological tumors, served as the reference for this study (procedures for H&E staining are provided in Supporting Information). Figure A compares the imaging results of H&E staining under bright-field imaging and CPD staining under TPF imaging at the same location. According to the pathological diagnosis, there are four categories: (NT) normal breast tissue, (I) pathological stage I BC tissue, (II) pathological stage II BC tissue, and (III) pathological stage III BC tissue. In normal breast tissue sections, minimal fluorescence emission was detected in the red channel, which was consistent with the low lipid accumulation level in normal tissues. In contrast, pathological stage I BC tissue sections showed slight fluorescence intensity after the CPD-based probe staining, while stage II sections displayed a progressively stronger fluorescence signal, and the strongest fluorescence was observed in stage III BC tissue sections. It should be noted that the increase in TPF fluorescence intensity of the CPD probe was positively correlated with the elevation of lipid metabolic activity in BC tissues, but there was no direct causal relationship between them. The TPF fluorescence intensity is a comprehensive reflection of multiple factors, including lipid content (number and size of lipid droplets), lipid droplet distribution, and tissue structural density, rather than a direct indicator of lipid metabolic activity. Therefore, the enhanced TPF signal in this study only represents the lipid-enriched phenotype of BC tissues caused by abnormal lipid metabolism, which is an external manifestation of elevated lipid metabolic activity, but cannot be directly equated with the comprehensive lipid metabolic state. Figure B shows the TPF images of normal tissues and various cancer stages, demonstrating consistent trends in fluorescence intensity within each stage. The average fluorescence intensity of TPF images from 417 frozen sections was calculated by converting all images into grayscale maps and extracting their pixel values. Figure C shows the detailed mean fluorescence intensity for each image, revealing significant variations across different stages. As illustrated in the violin plot in Figure D, 101 normal tissue sections showed weak fluorescence emission in the red channel, with an average intensity of approximately 0.01. The average intensity for stage I BC sections increased 3-fold to about 0.03, while the intensities for stages II and III rose approximately 90-fold and 140-fold, respectively. These statistical results suggest that CPD can effectively assist in the initial diagnosis of clinicopathological sections by highlighting the differences in lipid accumulation level between normal and cancerous breast tissues. The significant difference in TPF fluorescence intensity among different pathological stages is the direct reflection of the difference in lipid-enriched phenotype, which indirectly reflects the abnormal degree of lipid metabolism in BC tissues at different stages.

4.

4

Open in a new tab

(A) Frozen tissue sections at different clinicopathologic stages after H&E or CPD staining. From the top to bottom panels: bright-field images for H&E-stained tissue sections, bright-field images of tissue sections, TPF images of tissue sections after they were stained with CPD, and the merged images. Scale bar: 50 μm. (B) Five randomly selected TPF images of the frozen tissue sections at different clinicopathologic stages. Scale bar: 50 μm. λex = 790 nm. (C) Histogram of the averaged fluorescence intensity of each image. The total number of images is 417. (D) Violin plot of the averaged fluorescence intensity. *: p ⩽ 0.05, **: p ⩽ 0.01, ***: p ⩽ 0.001. ****: p ⩽ 0.0001.

3.4. Staging Tumor Grades by Pixel Point Distribution Maps and Machine Learning

Relying solely on imaging brightness data to stage breast tissues is easily affected by fluctuations in laser power, light path coupling efficiency, imaging camera temperature, and other factors. These introduce many unpredictable elements that impact the accuracy of the existing classification model. Over the past decade, machine learning disease diagnosis has attracted significant attention in various fields of medicine, biology, and chemistry, and has evolved into a generalized and modular workflow: initial screening of multiple biomarkers for a specific disease is performed, and then data is collected from these biomarkers and subsequently trained using appropriate machine learning algorithms to enable direct screening, classification, or diagnosis of the disease. To accurately evaluate the pathological stage of BC for rapid intraoperative pathology diagnosis, we obtained and compared pixel distribution maps across various stages of evolution, expecting to reveal significant differences with higher accuracy.

Figure S7A–D displays the original pixel distribution maps for four stages. The horizontal axis indicates the pixel gray value, while the vertical axis shows the number of pixels within that interval of gray values. These differences are apparent in the pixel distribution maps. As the disease progresses, more pixels are found in higher gray value intervals. To standardize the process, the pixel distributions of all images were unified to 10, and the pixel distribution maps were redrawn (Figure A–D).

5.

5

Open in a new tab

(A–D) Pixel distribution for each classification after normalization. (E) Distribution points for the four classifications after t-SNE dimensionality reduction. (F) Confusion matrix for test sets of random forest classifiers. (G) Heatmap of grouped clustering for four pathological stages.

The advent of machine learning has improved the ability to tackle classification and regression challenges in large image data sets. Predictive class algorithms based on supervised learning, such as support vector machines, Bayesian classification, decision trees, K-nearest neighbors, and neural networks, rely on external variables to generate interpretable models for predicting new samples. When combined with images, these methods aid in resolving complex classification issues in biological imaging. , Here, a machine learning algorithm was then applied to analyze the distribution of pixel values within the same interval across the four stages. Figure E shows the results of a t-SNE dimensionality reduction, clearly distinguishing four stages. Subsequently, the pixel distribution matrices for four stages were trained using the Random Forest classifier to accurately predict the stages of unknown specimens. The confusion matrices of the test set from the Random Forest classifier are shown in Figure F. The accuracy rates were 100% for normal breast tissue, 95.455% for stage I, 95% for stage II, and 95% for stage III BC tissue, all exceeding 95%, indicating effective classification for predicting the staging of unknown samples. Figure G illustrates the results of different grouping clustering heatmaps, where normal breast tissue and pathological stages I, II, and III are clustered into two major classes, with further subclass clustering observed. These results demonstrate that machine learning algorithms can effectively differentiate the pixel distributions across the four BC stages, enabling rapid intraoperative pathology diagnosis of unknown pathological samples. In short, the distribution of image pixels can be used to accurately classify and predict tissue sections across different stages.

3.5. CNNs Aid in Accurate Staging

Deep learning applications in the biomedical field are profoundly changing the paradigm of disease diagnosis, treatment, and research. As an important branch of deep learning, convolutional neural network (CNN) has shown outstanding advantages in medical image analysis by virtue of its local perception, weight sharing, and hierarchical feature extraction capabilities: for example, by training end-to-end models, CNN can automatically identify tumor foci in X-rays, computed tomographies (CTs), and magnetic resonance imagings (MRIs), accurately segment the tissue boundaries, or classify the subtypes of cancer cells in pathology slices, with an accuracy rate of 90% or more. With an accuracy rate of 90% or more, some scenarios surpass the level of human experts, and the breakthroughs in CNN technology are continuing to promote clinical translation in this field. − CNN was utilized for image classification. The structural composition of the CNN, which demonstrates a four-class classification model, is illustrated in Figure S8. To optimize model parameters and enhance CNN classification accuracy, four sets of neuron parameters were tested. The classification performance of these sets was evaluated based on the accuracy of the training and test sets, as shown in Figure A. Group d achieved the best results, with a training set accuracy of nearly 1, a test set accuracy above 0.9, and a more concentrated distribution of results across ten random tests. Figure B indicates that Group d also achieved the lowest cross-entropy loss, with the training set losses nearing 0 and the test set losses at approximately 0.2. Consequently, Group d parameters were selected for subsequent modeling. This optimized model presents changes in accuracy and cross-entropy loss in Figure C–D. Over the course of 30 iterations, the training set loss reduced from 1.4 to 0, while the test set loss decreased from 1.38 to 0.2. The training set accuracy improved from 0.26 to 1, and the test set accuracy rose from 0.21 to 0.95, indicating outstanding classification performance. Further, the accuracy of each category is illustrated in the confusion matrices. Figure E shows the training set, achieving 100% accuracy across all four stages, and Figure F shows the test set, with accuracies of 100%, 94.44%, 96%, and 90.91% for each stage, respectively. This high classification accuracy demonstrates the CNN model’s effectiveness in distinguishing among the four stages of BC, providing a rapid diagnostic tool for intraoperative pathology in unknown samples.

6.

6

Open in a new tab

(A) Scatterplot of the change in accuracy for 4 sets of parameters randomized 10 times. (B) Scatterplot of the change in loss values for 4 sets of parameters randomized 10 times. (C–D) Accuracy and loss curves for 30 iterations of four classification models. Black and red dots stand for the training set and the test set. (E, F) Four confusion matrices for classification models. (E) Training set. (F) Test set.

4. Conclusions

A machine learning-powered imaging technique has been developed for the rapid intraoperative pathological diagnosis of BC pathological stages by characterizing abnormalities in lipid metabolism with a TPF CPD-based probe. This method enables rapid diagnosis based on TPF images of intraoperative frozen sections that undergo rapid staining, offering several advantages over traditional H&E staining. It is simpler and faster, requiring only a 5 min incubation time, and eliminates the need for complex manipulations. Most importantly, this method reduces the reliance on specialized diagnostic procedures by professional pathologists, simplifying the entire testing process and making it accessible without specialized knowledge. TPF images were obtained across four pathological stages. Compared with the latest research advances in intraoperative rapid pathological diagnosis and the application of machine learning in pathological diagnosis, − we classified the data set using pixel distribution maps, machine learning, and convolutional neural network models, achieving favorable classification results. Grouped clustered heatmaps have effectively differentiated normal breast tissue from cancer cells across different staged sections. The rapid intraoperative pathological diagnosis of tumors is invaluable for early cancer detection and treatment evaluation. Ultimately, CPD combined with TPF microscopy represents a significant advancement in detecting abnormalities in lipid metabolism in pathological frozen sections, thereby enhancing the role of fluorescent probes in the pathological diagnosis of related diseases.

Supplementary Material

im6c00025_si_001.pdf (1,020.2KB, pdf)

Acknowledgments

We thank Prof. Bai Yang, Jilin University, for his helpful discussion on CPD synthesis. This work was supported by the National Natural Science Foundation of China NSFC (Nos. 22373041 and 22173035), the National Key Research and Development Program of China (2023YFC3706002), the Fundamental Research Funds for the Central Universities, Science and Technology Development Program Projects of Jilin Province (20220101046JC), and the Innovative Research Project of the State Key Laboratory of Supramolecular Structure and Materials.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/cbmi.6c00025.

  • Reagents and chemicals, apparatus, synthesis of cpd, preparation of O/W emulsions, cell experiments, raw pixel processing, pathological experiments, TPF Fluorescence Imaging, Oil Red O-stained images, composition of convolutional neural networks for classification models, and tables for age and pathological parameters of patients, elemental compositions of the CPD, and XPS data analyses of the C 1s, N 1s, and O 1s spectra of the CPD (PDF)

Z.Y.: Conceptualization, data curation, formal analysis, investigation, methodology, validation, writingoriginal draft. J.W.: Data curation, methodology. Y.L.: Data curation, methodology. W.X.: Funding acquisition, methodology, supervision. B.H.: Supervision, resource, investigation. S.X.: Conceptualization, funding acquisition, methodology, project administration, supervision, writing–review & editing.

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. This study was approved by an institutional ethics committee and followed the tenets of the Declaration of Helsinki (Ethics approval number: 2022–449).

The authors declare no competing financial interest.

References

  1. Lu Y., Zhang J., Liu X., Zhang Z., Li W., Zhou X., Li R.. Prediction of breast cancer metastasis by deep learning pathology. IET Image Process. 2023;17(2):533–543. doi: 10.1049/ipr2.12652. [DOI] [Google Scholar]
  2. Veta M., Pluim J. P., Van Diest P. J., Viergever M. A.. Breast cancer histopathology image analysis: A review. IEEE Trans. Biomed. Eng. 2014;61(5):1400–1411. doi: 10.1109/TBME.2014.2303852. [DOI] [PubMed] [Google Scholar]
  3. Pan X., Lu Y., Lan R., Liu Z., Qin Z., Wang H.. et al. Mitosis detection techniques in H&E stained breast cancer pathological images: A comprehensive review. Comput. Electr. Eng. 2021;91:107038. doi: 10.1016/j.compeleceng.2021.107038. [DOI] [Google Scholar]
  4. Wang P., Hu X., Li Y., Liu Q., Zhu X.. Automatic cell nuclei segmentation and classification of breast cancer histopathology images. Signal Process. 2016;122:1–13. doi: 10.1016/j.sigpro.2015.11.011. [DOI] [Google Scholar]
  5. de Haan K., Zhang Y., Zuckerman J. E., Liu T., Sisk A. E., Diaz M. F. P., Jen K.-Y., Nobori A., Liou S., Zhang S., Riahi R., Rivenson Y., Wallace W. D., Ozcan A.. Deep learning-based transformation of H&E stained tissues into special stains. Nat. Commun. 2021;12(3):4884. doi: 10.1038/s41467-021-25221-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Cheng H., Wang M., Su J., Li Y., Long J., Chu J.. et al. Lipid metabolism and cancer. Life. 2022;12(6):784. doi: 10.3390/life12060784. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Liu H., Zhang Z., Song L., Gao J., Liu Y.. Lipid metabolism of cancer stem cells. Oncol. Lett. 2022;23(4):119. doi: 10.3892/ol.2022.13239. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Poulose N., Amoroso F., Steele R. E., Singh R., Ong C. W., Mills I. G.. Genetics of lipid metabolism in prostate cancer. Nat. Genet. 2018;50(2):169–171. doi: 10.1038/s41588-017-0037-0. [DOI] [PubMed] [Google Scholar]
  9. Liu X., Zhang P., Xu J., Lv G., Li Y.. Lipid metabolism in tumor microenvironment: novel therapeutic targets. Cancer Cell Int. 2022;22(1):224. doi: 10.1186/s12935-022-02645-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Lou J., Ancajas C. F., Zhou Y., Lane N. S., Reynolds T. B., Best M. D.. Probing Glycerolipid Metabolism using a Caged Clickable Glycerol-3-Phosphate Probe. ChemBioChem. 2024;25(18):e202300853. doi: 10.1002/cbic.202300853. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Terry A. R., Hay N.. Emerging targets in lipid metabolism for cancer therapy. Trends Pharmacol. Sci. 2024;45(6):537–551. doi: 10.1016/j.tips.2024.04.007. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Majchrzak M., Stojanović O., Ajjaji D., Ben M’barek K., M’barek K. B., Omrane M., Thiam A. R.. Perilipin membrane integration determines lipid droplet heterogeneity in differentiating adipocytes. Cell Rep. 2024;43(4):114093. doi: 10.1016/j.celrep.2024.114093. [DOI] [PubMed] [Google Scholar]
  13. Matsushita Y., Nakagawa H., Koike K.. Lipid metabolism in oncology: why it matters, how to research, and how to treat. Cancers. 2021;13(3):474. doi: 10.3390/cancers13030474. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Jin H. R., Wang J., Wang Z. J., Xi M. J., Xia B. H., Deng K., Yang J. L.. Lipid metabolic reprogramming in tumor microenvironment: from mechanisms to therapeutics. J. Hematol. Oncol. 2023;16(1):103. doi: 10.1186/s13045-023-01498-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Walther T. C., Farese Jr R. V.. Lipid droplets and cellular lipid metabolism. Annu. Rev. Biochem. 2012;81(1):687–714. doi: 10.1146/annurev-biochem-061009-102430. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Wang N., Wang J., Wang P., Ji N., Yue S.. Label-free Raman spectromicroscopy unravels the relationship between MGMT methylation and intracellular lipid accumulation in glioblastoma. Anal. Chem. 2023;95(31):11567–11571. doi: 10.1021/acs.analchem.3c00967. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Zeng J., Wu G., Jia H., Liu C., Liu C., Zhang Y.. et al. Stimulated Raman scattering microscopy reveals aberrant triglyceride accumulation in lymphatic metastasis of papillary thyroid carcinoma. Anal. Chem. 2024;96(15):5735–5740. doi: 10.1021/acs.analchem.4c00306. [DOI] [PubMed] [Google Scholar]
  18. Biswas S., Baruah M., Shil A., Sarkar S., Ali M., Samanta A., Bhuniya S.. Polarity-driven two-photon fluorescent probe for monitoring the perturbation in lipid droplet levels during mitochondrial dysfunction and acute pancreatitis. ACS Sens. 2023;8(10):3793–3803. doi: 10.1021/acssensors.3c01245. [DOI] [PubMed] [Google Scholar]
  19. Han J., Lee H. W., Chen Y., Li H., Kim H. M., Yoon J.. Observing hepatic steatosis with a commercially viable two-photon fluorogenic probe. Mater. Chem. Front. 2022;6(5):553–560. doi: 10.1039/D1QM01665B. [DOI] [Google Scholar]
  20. Lu S., Xiao G., Sui L., Feng T., Yong X., Zhu S.. et al. Piezochromic carbon dots with two-photon fluorescence. Angew. Chem. 2017;129(22):6283–6287. doi: 10.1002/ange.201700757. [DOI] [PubMed] [Google Scholar]
  21. Wu T., Duan L., Yang J., Zhou Y.. A multidimensional quantitative fluorescent probe for high-fidelity monitoring of the polarity change in lysosomes during autophagy. Sens. Actuators, B. 2023;393:134142. doi: 10.1016/j.snb.2023.134142. [DOI] [Google Scholar]
  22. Xu H., Zhang H., Liu G., Kong L., Zhu X., Tian X.. et al. Coumarin-based fluorescent probes for super-resolution and dynamic tracking of lipid droplets. Anal. Chem. 2019;91(1):977–982. doi: 10.1021/acs.analchem.8b04079. [DOI] [PubMed] [Google Scholar]
  23. Yang C., Jiang C., Yang M., Bai Q., Zhen Y., Zhang Y.. et al. NAD (P) H activated fluorescent probe for rapid intraoperative pathological diagnosis and tumor histological grading. Chem. Biomed. Imaging. 2023;1(8):738–749. doi: 10.1021/cbmi.3c00076. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Li D., Deng X., Xu Z., Wang D., Xu G., Zhang P.. et al. Molecular engineering of NIR-II AIE luminogen excited at 1700 nm for ultradeep intravital brain two-photon fluorescence imaging. Adv. Funct. Mater. 2023;33(46):2303967. doi: 10.1002/adfm.202303967. [DOI] [Google Scholar]
  25. Zheng Z., Yang Y., Wang P., Gou X., Gong J., Wu X.. et al. A bright two-photon lipid droplets probe with viscosity-enhanced solvatochromic emission for visualizing lipid metabolic disorders in deep tissues. Adv. Funct. Mater. 2023;33(35):2303627. doi: 10.1002/adfm.202303627. [DOI] [Google Scholar]
  26. Ma Z., Fu Z., Hou B., Zeng X., Wang J., Tan Y., Jiang Y., Xu N., Tan C.. Two-Photon Light-Activatable Fluorophores for Organelle Imaging in Living Cells and Tissue-Level Imaging. J. Am. Chem. Soc. 2025;147:37005–37011. doi: 10.1021/jacs.5c14442. [DOI] [PubMed] [Google Scholar]
  27. Huo Z., Cao X., Sun D., Xu W., Yang B., Xu S.. Carbonized polymer dot probe for two-photon fluorescence imaging of lipid droplets in living cells and tissues. ACS Sens. 2023;8(5):1939–1949. doi: 10.1021/acssensors.2c02678. [DOI] [PubMed] [Google Scholar]
  28. Huo Z., Yu Z., Xu W., Xu S.. Super-Resolution microscopic imaging of lipid droplets in living cells via carbonized polymer dot-based polarity-responsive nanoprobe. ACS Meas. Sci. Au. 2024;4:593–598. doi: 10.1021/acsmeasuresciau.4c00049. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Yu Z., Cao X., Huo Z., Xu W., Xu S.. In Situ Two-photon fluorescence spectroscopy to precisely reveal intracellular lipid droplet microenvironment via a carbonized polymer dot nanosensor. Sens. Actuators, B. 2025;424:136900. doi: 10.1016/j.snb.2024.136900. [DOI] [Google Scholar]
  30. Yu Z., Huo Z., Xu W., Xu S.. Multicolor lipid droplet-targeted carbonized polymer dots with luminescence regulation for bioimaging applications. ACS Appl. Nano Mater. 2024;7:9068–9076. doi: 10.1021/acsanm.4c00545. [DOI] [Google Scholar]
  31. Abramczyk H., Surmacki J., Kopeć M., Olejnik A. K., Lubecka-Pietruszewska K., Fabianowska-Majewska K.. The role of lipid droplets and adipocytes in cancer. Raman imaging of cell cultures: MCF10A, MCF7, and MDA-MB-231 compared to adipocytes in cancerous human breast tissue. Analyst. 2015;140(7):2224–2235. doi: 10.1039/C4AN01875C. [DOI] [PubMed] [Google Scholar]
  32. Ghosh C., Nandi S., Bhattacharyya K.. Probing micro-environment of lipid droplets in a live breast cell: MCF7 and MCF10A. Chem. Phys. Lett. 2017;670:27–31. doi: 10.1016/j.cplett.2016.12.068. [DOI] [Google Scholar]
  33. Greenberg A. S., Coleman R. A.. Expanding roles for lipid droplets. Trends Endocrinol. Metab. 2011;22(6):195–196. doi: 10.1016/j.tem.2011.04.002. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Blucher C., Stadler S. C.. Obesity and Breast Cancer: Current Insights on the Role of Fatty Acids and Lipid Metabolism in Promoting Breast Cancer Growth and Progression. Front Endocrinol (Lausanne) 2017;8:293. doi: 10.3389/fendo.2017.00293. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Zheng J.-S., Hu X.-J., Zhao Y.-M., Yang J., Li D.. Intake of fish and marine n-3 polyunsaturated fatty acids and risk of breast cancer: meta-analysis of data from 21 independent prospective cohort studies. BMJ. 2013;346:f3706. doi: 10.1136/bmj.f3706. [DOI] [PubMed] [Google Scholar]
  36. Terry A. R., Nogueira V., Rho H., Ramakrishnan G., Li J., Kang S.. et al. CD36 maintains lipid homeostasis via selective uptake of monounsaturated fatty acids during matrix detachment and tumor progression. Cell Metab. 2023;35(11):2060–76. e9. doi: 10.1016/j.cmet.2023.09.012. [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Wang C., Shi M., Ji J., Cai Q., Zhao Q., Jiang J.. et al. Stearoyl-CoA desaturase 1 (SCD1) facilitates the growth and anti-ferroptosis of gastric cancer cells and predicts poor prognosis of gastric cancer. Aging. 2020;12(15):15374. doi: 10.18632/aging.103598. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Singh V., Chassaing B., Zhang L., Yeoh B. S., Xiao X., Kumar M.. et al. Microbiota-dependent hepatic lipogenesis mediated by stearoyl CoA desaturase 1 (SCD1) promotes metabolic syndrome in TLR5-deficient mice. Cell Metab. 2015;22(6):983–996. doi: 10.1016/j.cmet.2015.09.028. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Ascenzi F., De Vitis C., Maugeri-Saccà M., Napoli C., Ciliberto G., Mancini R.. SCD1, autophagy and cancer: implications for therapy. J. Exp. Clin. Cancer Res. 2021;40(1):265. doi: 10.1186/s13046-021-02067-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Zhang J., Song F., Zhao X., Jiang H., Wu X., Wang B.. et al. EGFR modulates monounsaturated fatty acid synthesis through phosphorylation of SCD1 in lung cancer. Mol. Cancer. 2017;16(1):127. doi: 10.1186/s12943-017-0704-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Ibrahim A., Gamble P., Jaroensri R., Abdelsamea M. M., Mermel C. H., Chen P.-H.C., Rakha E. A.. Artificial intelligence in digital breast pathology: techniques and applications. Breast. 2020;49:267–273. doi: 10.1016/j.breast.2019.12.007. [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. Huang L., Sun H., Sun L., Shi K., Chen Y., Ren X.. et al. Rapid, label-free histopathological diagnosis of liver cancer based on Raman spectroscopy and deep learning. Nat. Commun. 2023;14(1):48. doi: 10.1038/s41467-022-35696-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Liu Z., Chen L., Cheng H., Ao J., Xiong J., Liu X.. et al. Virtual formalin-fixed and paraffin-embedded staining of fresh brain tissue via stimulated Raman CycleGAN model. Sci. Adv. 2024;10(13):eadn3426. doi: 10.1126/sciadv.adn3426. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Djavanshir G. R., Chen X., Yang W.. A review of artificial intelligence’s neural networks (deep learning) applications in medical diagnosis and prediction. IT Prof. 2021;23(3):58–62. doi: 10.1109/MITP.2021.3073665. [DOI] [Google Scholar]
  45. Ibrahim A., Gamble P., Jaroensri R.. et al. Artificial intelligence in digital breast pathology: techniques and applications. Breast. 2020;49:109–121. doi: 10.1016/j.breast.2019.12.007. [DOI] [PMC free article] [PubMed] [Google Scholar]
  46. Djavanshir G. R., Chen X., Yang W.. A review of artificial intelligence’s neural networks (deep learning) applications in medical diagnosis and prediction. IT Prof. 2021;25(2):45–52. doi: 10.1109/mitp.2021.3073665. [DOI] [Google Scholar]
  47. Wu S., Yue M., Zhang J., Li X., Li Z.. et al. The role of artificial intelligence in accurate interpretation of her2 immunohistochemical scores 0 and 1+ in breast cancer. Mod. Pathol. 2023;36(3):100054. doi: 10.1016/j.modpat.2022.100054. [DOI] [PubMed] [Google Scholar]
  48. Kim M., Hong S., Yankeelov T. E., Yeh H., Liu Y.. Deep learning-based classification of breast cancer cells using transmembrane receptor dynamics. Bioinformatics. 2021;38(1):243–249. doi: 10.1093/bioinformatics/btab581. [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Munshi R. M., Cascone L., Alturki N., Saidani O., Alshardan A., Umer M.. A novel approach for breast cancer detection using optimized ensemble learning framework and XAI. Image Vis. Comput. 2024;142:104910. doi: 10.1016/j.imavis.2024.104910. [DOI] [Google Scholar]
  50. Meng F., Liu Q., Zhu W., He J., Gao P., Gao L., Yu X.. Intraoperative pathological diagnosis for breast cancer with a polarity fluorescent probe targeting lipid droplets. Sens. Actuators, B. 2025;443:138305. doi: 10.1016/j.snb.2025.138305. [DOI] [Google Scholar]
  51. Ropelewska E., Rutkowski K.. The classification of peaches at different ripening stages using machine learning models based on texture parameters of flesh images. Agriculture. 2023;13(2):498. doi: 10.3390/agriculture13020498. [DOI] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

im6c00025_si_001.pdf (1,020.2KB, pdf)

Articles from Chemical & Biomedical Imaging are provided here courtesy of American Chemical Society

RESOURCES