Predicting Helicobacter pylori Antibiotic Resistance from Routine Hematoxylin and Eosin Histopathology Using a Vision-Language Model-Guided Foundation Model Framework Authors: Siping Xiong, Shuguang Liu, Wei Zhang, Degui Liao, Tian Tang, Chao Zeng, Yimin Guo Published: 2026 ABSTRACT AbstractBackgroundHelicobacter pylori (H. pylori) eradication is increasingly compromised by antibiotic resistance, particularly to clarithromycin (CLA) and fluoroquinolones (FQs). Culture-based and molecular susceptibility testing remain resource-intensive and inaccessible in many settings. Whether routine hematoxylin and eosin (H&E) histopathology encodes indirect signatures of resistance has not been explored.Materials and MethodsWe developed a weakly supervised deep learning framework to predict phenotypic H. pylori resistance directly from gastric biopsy whole-slide images (WSIs). Patches were filtered using vision-language models (VLMs; Qwen3-VL and MedGemma) performing two zero-shot classification tasks—grading patch image quality and identifying gastric surface or gland-neck epithelium—based solely on visible histo-morphological features. Retained patches were encoded with VIRCHOW2, a histopathology-specific vision foundation model (ViT-H/14, pretrained on 3.1 million WSIs), and aggregated via cross-attention for slide-level prediction. Polymerase chain reaction (PCR)-confirmed resistance phenotypes served as ground truth in a multicenter cohort of 755 patients.ResultsIn the independent test set (n = 151), the framework achieved an area under the receiver operating characteristic curve (AUC) of 0.903 for CLA resistance and 0.959 for FQs resistance. At the default threshold (0.5), sensitivity was 85.2% for CLA and 98.1% for FQs, supporting use as a screening tool. Threshold optimization by the Youden index improved specificity to 95.9% for CLA and 93.9% for FQs, enabling rule-in risk stratification. Decision curve analysis confirmed net clinical benefit across a broad range of threshold probabilities.ConclusionRoutine H&E-stained gastric biopsies contain probabilistic morphological correlates of H. pylori resistance phenotypes that deep learning can exploit. This approach provides complementary decision support for resistance risk stratification, particularly in resource-limited settings, pending prospective validation.Keywords: Helicobacter pylori, antibiotic resistance, deep learning, computational pathology, vision foundation model, hematoxylin and eosin staining, whole-slide imaging Introduction Helicobacter pylori (H. pylori) infection remains one of the most prevalent chronic bacterial infections worldwide, affecting approximately half of the global population, and is a well-established cause of chronic gastritis, peptic ulcer disease, and gastric adenocarcinoma.1 Eradication therapy significantly reduces disease progression and gastric cancer risk; however, its effectiveness has been increasingly compromised by rising antibiotic resistance, particularly to clarithromycin (CLA), metronidazole, and fluoroquinolones (FQs).2,3 CLA-resistant H. pylori has been designated a high-priority pathogen by the World Health Organization (WHO).4 In many regions, primary resistance rates now exceed thresholds at which standard empirical therapies are no longer recommended, underscoring the urgent need for individualized treatment strategies.5,6 Current approaches for determining H. pylori antimicrobial susceptibility rely on culture-based phenotypic testing or molecular assays targeting resistance-associated mutations—point mutations in the 23S rRNA gene for CLA resistance or gyrA mutations for FQs resistance.7,8 While reliable, their clinical adoption is limited by technical complexity, cost, prolonged turnaround time, and restricted availability, particularly in low-resource settings. Antibiotic selection therefore often remains empirical, contributing to repeated treatment failures and further resistance evolution.9 Routine histopathological examination using hematoxylin and eosin (H&E) staining is the standard diagnostic modality for H. pylori-associated gastritis. Beyond confirming infection, H&E histopathology provides rich information on host tissue responses—inflammatory activity (neutrophil infiltration), chronic infiltrates, lymphoid aggregates, glandular atrophy, and intestinal metaplasia—as codified by the Updated Sydney System.10–12 Whether such morphological patterns encode indirect signatures associated with resistance phenotypes has not been systematically investigated. Antibiotic resistance in H. pylori arises primarily from bacterial genetic alterations not directly visible on routine histology. However, resistant strains often produce refractory infections that may exhibit altered colonization behavior, enhanced survival under therapeutic pressure, or differential induction of host inflammatory responses compared with susceptible strains.13 These downstream tissue-level manifestations raise the possibility that resistance-associated phenotypes may be indirectly reflected in H&E-stained sections. We hypothesized, moreover, that the two resistance phenotypes may leave partially distinct tissue-level footprints: clarithromycin resistance, frequently linked to persistent, treatment-refractory infection, may be accompanied by more pronounced chronic and atrophic mucosal remodeling, whereas fluoroquinolone resistance may correlate with different spatial patterns of glandular and peri-glandular inflammation—hypotheses that motivate a spatially resolved morphological analysis. Recent advances in deep learning-based computational pathology have demonstrated that vision transformers can extract high-dimensional morphological representations from whole-slide images (WSIs) to predict clinically relevant molecular or genetic states—including microsatellite instability, BRAF/KRAS mutations, and treatment response—directly from H&E histology.14–17 However, general-purpose vision models pretrained on natural images transfer poorly to histopathology, because they inadequately capture the stain-dependent, tissue-specific morphology of tissue sections; this limitation has motivated histopathology-specific foundation models pretrained on large-scale WSIs. Applying such domain-adapted methodology to investigate associations between gastric histomorphology and H. pylori antibiotic resistance represents a novel and clinically motivated extension. In this study, we evaluate the feasibility of predicting H. pylori antibiotic resistance from routine H&E-stained gastric biopsy WSIs using a weakly supervised framework. Tissue regions are first filtered using vision-language models (VLMs) based solely on visible histo-morphological features, then encoded using a histopathology-specific vision foundation model (VIRCHOW2) and aggregated via cross-attention. Our objective is not to replace molecular or culture-based assays, but to assess whether histology-based models may provide complementary information for resistance risk stratification in settings where access to dedicated resistance testing is limited. Materials and Methods Study Population and Dataset This multicenter retrospective study included 755 gastric mucosal biopsy cases from two tertiary institutions. The primary cohort comprised 674 cases from the Eighth Affiliated Hospital of Sun Yat-sen University (SYSU8H); the secondary cohort included 81 cases from the Second Affiliated Hospital of Guangzhou Medical University (GMU2H). The cohort included both treatment-naïve patients and those with prior failed eradication, reflecting real-world clinical practice. The study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Boards of both institutions (protocol code D2025R052). Written informed consent was waived owing to the retrospective nature of the study. Resistance was determined from paraffin-embedded gastric tissue using amplification refractory mutation system polymerase chain reaction (ARMS-PCR) assays (Xinji Yongkang Biotech, Beijing, China). CLA resistance was defined by mutations at three 23S rRNA loci (A2142G, A2143G, A2142C); FQs resistance by mutations at six gyrA loci (A260T, T261G/C261G, T261A/C261A, G271A, G271T, A272G). H&E-stained slides were digitized using a PANNORAMIC 1000 scanner (3DHISTECH, Budapest, Hungary) at 10× magnification (~0.5 µm/pixel). WSIs were tessellated into non-overlapping 512 × 512-pixel patches; background was removed by Otsu thresholding. Partitioning was performed at the patient level. The SYSU8H cohort was divided into a training set (n = 453) and an internal validation set (n = 151). An independent test set (n = 151) was formed by combining the remaining SYSU8H cases (n = 70) with all GMU2H cases (n = 81). Cohort characteristics and data-split proportions are summarized in Table 1 and Figure 1a–d. Table 1Distribution of Tissue Samples and Clinicopathological Characteristics Across DatasetsParameterTrainingValidationTestSample size, n453151151Age, mean ± SD, years54.1 ± 11.753.1 ± 11.951.8 ± 12.0Sex, male, n (%)269 (59.4)88 (58.3)87 (57.6)CLA susceptibility, n (%) Sensitive299 (66.0)97 (64.2)97 (64.2) Resistant154 (34.0)54 (35.8)54 (35.8)FQs susceptibility, n (%) Sensitive315 (69.5)96 (63.6)99 (65.6) Resistant138 (30.5)55 (36.4)52 (34.4)Abbreviations: CLA, clarithromycin; FQs, fluoroquinolones; SD, standard deviation. Distribution of Tissue Samples and Clinicopathological Characteristics Across Datasets Abbreviations: CLA, clarithromycin; FQs, fluoroquinolones; SD, standard deviation. Stratified distribution of clinical characteristics. Sankey diagrams show the proportional flow of patient sex (left), CLA susceptibility (middle), and FQs susceptibility (right) for (a) the total dataset, (b) the training set, (c) the validation set, and (d) the independent test set. Uniform band widths confirm preserved class balance across splits. Deep Learning Framework The proposed pipeline comprises four sequential stages: (i) WSI preprocessing and tissue segmentation; (ii) VLM-based patch quality and region filtering; (iii) feature encoding using VIRCHOW2; and (iv) cross-attention aggregation and slide-level prediction (Figure 2a and b). Overview of the proposed framework for H. pylori antibiotic resistance prediction. (a) WSI preprocessing: tissue regions are isolated using contour extraction and tessellated into patches. VLM-based quality filtering (Qwen3-VL and MedGemma) retains epithelially relevant patches, followed by Macenko stain normalization. (b) Slide-level prediction: patches are encoded by VIRCHOW2 (ViT-H/14); embeddings are aggregated via cross-attention to generate a slide-level representation for CLA and FQs resistance prediction. WSI Preprocessing and Patch Extraction Given multicenter staining variability, an automated contour extraction procedure was implemented with OpenCV (version 4.8; open source). WSIs were converted to grayscale and thresholded by the Otsu method to generate binary tissue masks excluding background (mean pixel intensity > 240). Tissue was tessellated into non-overlapping 512 × 512-pixel patches at 10× magnification. Macenko stain normalization was applied to all retained patches to reduce inter-center color variation. The 10× magnification was selected deliberately: it preserves the glandular architecture, inflammatory patterning, and broader tissue-level context that are most relevant to host-response morphology, while avoiding the substantially larger patch counts and computational overhead incurred at higher magnifications, which emphasize single-cell detail over the spatial context of interest. VLM-Based Patch Filtering An automated patch-level filtering step using VLMs was applied to remove non-informative regions. We evaluated Qwen3-VL-8B-Instruct18 and MedGemma-1.5–4B-IT19 using identical prompts on five randomly selected WSIs (~1000 patches). Each patch was assessed independently for overall image quality, binary classification of whether the patch primarily contained gastric surface and/or neck epithelium (TRUE/FALSE), and classification confidence. The prompt explicitly prohibited inference from clinical context, anatomical location, or prior knowledge unless directly supported by visible morphology. Annotations were reviewed by an experienced gastrointestinal pathologist. The complete prompt content supplied to both VLMs was as follows: You are an experienced gastrointestinal pathologist assisting with histopathology research data annotation. You will analyze 10× H&E-stained gastric mucosal image patches. All assessments must be based strictly on the histomorphological features visible in the image. Do not infer findings from clinical information, anatomical location, or prior knowledge unless they are directly supported by the observed morphology. Your tasks: 1. Assess the overall image quality for histopathological interpretation. 2. Determine whether the image primarily contains gastric surface epithelium and/or gastric neck epithelium. Classification Criteria: (TRUE) The image is predominantly composed of gastric surface (foveolar) epithelium and/or gastric neck epithelium. Typical features may include orderly columnar epithelial cells, preserved epithelial polarity, gastric pits (foveolae), and densely packed mucous cells characteristic of the neck region. (FALSE) The image is predominantly composed of deeper gastric glands, lamina propria, muscularis mucosae, background or non-tissue regions, or any region in which epithelial structures are not the dominant component. Rules: Base every decision solely on the visible histo-morphological features. If interpretation is limited by insufficient tissue, poor section orientation, staining artifacts, folding, blur, or other technical issues, this should be reflected in both the image quality and confidence scores. Avoid making overly confident judgments when the morphology is ambiguous. Always provide a binary classification (TRUE or FALSE), an image quality assessment, and a confidence score for your classification. Patches with insufficient image quality or classified as FALSE for relevant gastric surface/neck epithelium were excluded. Qwen3-VL-8B-Instruct demonstrated more consistent identification of relevant epithelial regions and was selected for dataset-wide filtering. Approximately 30% of patches were excluded across the dataset due to insufficient quality or absence of relevant gastric epithelial content. Feature Encoding with VIRCHOW2 Patch-level features were extracted using VIRCHOW2,20 a vision transformer (ViT-H/14) pretrained via self-supervised learning on approximately 3.1 million histopathology WSIs. Each normalized patch (224 × 224 pixels) was encoded into a 1280-dimensional embedding. The self-attention mechanism in VIRCHOW2 captures long-range spatial relationships within patches, providing domain-adapted representations suited to gastric mucosal architecture. Cross-Attention Aggregation and Prediction Patch embeddings served as keys and values, while a learnable query vector computed attention weights via the scaled dot-product attention formulation. The aggregated slide-level representation was passed to a multilayer perceptron to estimate resistance probability for CLA and FQs. The network was trained end-to-end using binary cross-entropy loss with the AdamW optimizer (learning rate 1 × 10−4, cosine annealing). Training ran for 100 epochs (batch size 32, single NVIDIA RTX 6000 Pro graphics processing unit; NVIDIA, Santa Clara, CA, USA); the best checkpoint was selected by validation AUC. The model was implemented in Python (version 3.11) using PyTorch (version 2.2). Statistical Analysis Model discrimination was assessed with the AUC. Operating-point metrics—sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV)—were computed at the default threshold (0.5), at the Youden-optimal threshold, and at a high-sensitivity threshold selected for screening use. Calibration was evaluated using the expected calibration error (ECE) and reliability diagrams. Clinical utility was assessed by decision curve analysis (DCA). Predictive values were additionally computed across a range of plausible resistance prevalences. Continuous variables are expressed as mean ± standard deviation (SD). Because the resistance classes were imbalanced, the prevalence-independent AUC served as the primary discrimination metric and was complemented by prevalence-adjusted PPV and NPV and by the F1 score. Results Prediction of CLA Resistance The model demonstrated stable discriminative performance for CLA resistance across cohorts (Figure 3a). AUC values were 0.912 (training), 0.899 (validation), and 0.903 (independent test set), indicating preserved generalization. At the default threshold (0.5), test-set sensitivity was 85.2% and specificity was 68.0% (Table 2), reflecting a sensitivity-oriented operating point suitable for screening. Model performance evaluation. (a) Receiver operating characteristic (ROC) curves (left) and predicted probability density distributions (right) for CLA resistance prediction across training, validation, and test sets. (b) Corresponding plots for FQs resistance. Density plots show score distributions for resistant (blue) and sensitive (yellow) cases. Table 2Test-Set Performance at Selected Operating Points for CLA and FQs ResistanceTaskOperating PointThresholdSensitivity, %Specificity, %PPV, %NPV, %CLADefault0.50085.268.059.889.2CLAYouden index0.87573.495.990.986.6CLAHigh-sensitivity0.36396.350.552.196.1FQsDefault0.50098.162.657.998.4FQsYouden index0.88588.593.988.494.0FQsHigh-sensitivity0.57996.271.864.197.3Notes: Predictive values were calculated at observed test-set prevalence (CLA: 35.8%; FQs: 34.4%).Abbreviations: CLA, clarithromycin; FQs, fluoroquinolones; NPV, negative predictive value; PPV, positive predictive value. Test-Set Performance at Selected Operating Points for CLA and FQs Resistance Notes: Predictive values were calculated at observed test-set prevalence (CLA: 35.8%; FQs: 34.4%). Abbreviations: CLA, clarithromycin; FQs, fluoroquinolones; NPV, negative predictive value; PPV, positive predictive value. Threshold optimization using the Youden index (0.875) provided a specificity-oriented profile (specificity 95.9%, PPV 90.9%), appropriate for rule-in risk stratification. A high-sensitivity threshold (0.363) achieved sensitivity of 96.3% at reduced specificity. DCA (Figure 4A) demonstrated clear net benefit over treat-all and treat-none strategies across a broad range of threshold probabilities. Predictive values (PPV and NPV) across a range of plausible resistance prevalences are shown in Figure 4B, and the ROC curve annotated with the Youden-optimal and high-sensitivity operating points is shown in Figure 4C. The ECE was 0.1843 (Figure 4D), indicating effective ranking with moderate overestimation in intermediate probability ranges. Comprehensive performance evaluation for CLA resistance. (A) Decision curve analysis: net benefit of the CLA model (blue) versus treat-all (dashed red) and treat-none (dotted) strategies. (B) PPV and NPV across resistance prevalence levels; vertical dashed line marks observed prevalence (35.8%). (C) ROC curve (AUC 0.903) with Youden index (purple; threshold 0.875) and high-sensitivity (green; threshold 0.363) operating points. (D) Calibration curve with ECE = 0.1843; inset histogram shows predicted probability distribution. Prediction of FQs Resistance For FQs resistance, the framework achieved AUC values of 0.957 (training), 0.903 (validation), and 0.959 (test set; Figure 3b). At the default threshold (0.5), sensitivity was 98.1% and specificity was 62.6% (Table 2). The Youden-optimal threshold (0.885) markedly improved the balance (sensitivity 88.5%, specificity 93.9%, PPV 88.4%). DCA confirmed high clinical utility across nearly the full range of threshold probabilities (Figure 5A). Predictive values (PPV and NPV) across resistance prevalences are shown in Figure 5B, and the ROC curve annotated with the Youden-optimal and high-sensitivity operating points is shown in Figure 5C. The ECE was 0.2126 with a bimodal predicted probability distribution (Figure 5D), suggesting that recalibration may further enhance clinical deployment. Comprehensive performance evaluation for FQs resistance. (A) Decision curve analysis. (B) PPV and NPV across prevalence levels (observed: 34.4%). (C) ROC curve (AUC 0.959) with Youden index (purple; threshold 0.885) and high-sensitivity (green; threshold 0.579) operating points. (D) Calibration curve with ECE = 0.2126. Cross-Attention Importance Maps To examine the spatial distribution of predictive signals, patch-level attention weights were visualized (Figure 6a and b). For CLA prediction, high-importance patches were predominantly localized to surface epithelial regions with active inflammatory infiltration and epithelial injury. For FQs, attention was more spatially diffuse, extending across cervical mucous glands, proper glands, and peri-glandular inflammatory microenvironments. Additionally, a trend toward milder mucosal inflammation was observed in CLA- and/or FQs-resistant cases compared with susceptible cases. These patterns suggest that the two resistance phenotypes are associated with partially distinct morphological signal distributions, though associations remain probabilistic within the weakly supervised framework. Cross-attention-derived patch importance maps. (a) Importance weights for CLA resistance prediction. (b) Importance weights for FQs resistance prediction. Color coding reflects normalized attention scores: Orange (high importance, 0.75–1.00) to white (low importance, 0.00–0.25). Ablation Analysis To dissect the contributions of backbone architecture and patch filtering, we compared DINOv321 and VIRCHOW220 encoders with and without VLM filtering (Table 3). Replacing DINOv3 with VIRCHOW2 consistently improved AUC and F1 score across both antibiotics (CLA and FQs), reflecting the advantage of large-scale histopathology pretraining. VLM-based filtering further improved AUC when combined with VIRCHOW2, supporting the hypothesis that restricting analysis to morphologically relevant epithelial regions enhances signal-to-noise ratio. Together, domain-adapted foundation modeling and morphology-constrained patch selection contribute independently and synergistically to resistance prediction. Table 3Ablation: Effect of Encoder Architecture and VLM Filtering on Test-Set Performance for CLA and FQs Susceptibility PredictionEncoderVLM FilterCLAFQsAccAUCF1AccAUCF1DINOv3No67.674.617.076.282.253.9DINOv3Yes67.678.122.282.189.771.6VIRCHOW2No70.284.666.779.590.674.0VIRCHOW2Yes74.290.370.274.895.972.9Notes: Accuracy (Acc) and F1 score are reported at the default threshold (0.5).Abbreviations: AUC, area under the receiver operating characteristic curve; CLA, clarithromycin; FQs, fluoroquinolones; VLM, vision-language model. Ablation: Effect of Encoder Architecture and VLM Filtering on Test-Set Performance for CLA and FQs Susceptibility Prediction Notes: Accuracy (Acc) and F1 score are reported at the default threshold (0.5). Abbreviations: AUC, area under the receiver operating characteristic curve; CLA, clarithromycin; FQs, fluoroquinolones; VLM, vision-language model. Discussion This study demonstrates that routine H&E-stained gastric biopsy WSIs contain probabilistic morphological correlates of H. pylori antibiotic resistance phenotypes that a weakly supervised deep learning framework can exploit. The model achieved strong discriminative performance (AUC > 0.90 for both CLA and FQs) in an independent multicenter test set, without requiring any manual annotation beyond patient-level resistance labels. The biological basis for morphology-based resistance prediction requires careful interpretation. CLA resistance in H. pylori arises primarily from point mutations in the domain V of the 23S rRNA gene, and FQs resistance from gyrA mutations—neither directly visible on H&E. The model is therefore unlikely to be detecting resistance mechanisms themselves. Rather, the predictive signal most plausibly reflects downstream tissue-level correlates of resistant infection. Resistant strains, which persist under physiological or therapeutic pressure, may exhibit higher bacterial load or induce distinct micro-anatomical inflammatory patterns compared with susceptible strains—characterized by sustained neutrophilic activity, specific spatial distributions of lymphoid aggregates, or subtle epithelial remodeling.13 This paradigm is consistent with prior computational pathology studies demonstrating that deep learning can infer molecular states from histology even without direct morphological correlates.14,15 From a clinical perspective, the proposed approach is intended as a flexible, complementary decision-support tool rather than a replacement for established susceptibility testing. As demonstrated by DCA, the model can be tuned for high specificity (Youden threshold) to identify patients likely harboring resistance for rule-in decision support, or for high sensitivity (lower threshold) as a screening tool in resource-limited settings. In such settings—where culture or PCR-based susceptibility testing is unavailable—a positive screen could prioritize limited testing resources for patients most likely to fail empirical therapy. Limitations Several limitations warrant consideration. First, although 755 cases represent a substantial cohort for a histopathology-based analysis, the dataset size is modest relative to foundation model complexity; larger, more diverse datasets are needed to confirm robustness. Second, a critical potential confounder is treatment history. Resistance in H. pylori is frequently secondary—acquired following failed eradication. Patients with resistant strains may exhibit histological features of “refractory gastritis” (more severe atrophy, intestinal metaplasia, or inflammatory scarring) independent of the resistance genotype itself. The model may therefore partly learn to identify histological signatures of treatment failure rather than resistance per se. While this distinction limits pure biological specificity, it does not diminish clinical utility as a risk stratification tool: identifying patients with “refractory-like” histology may still guide the decision to pursue molecular testing before prescribing alternative regimens. Third, despite the inclusion of an independent multicenter test set, the retrospective design introduces potential selection biases. Variability in tissue processing, staining, and slide scanning across institutions may affect generalizability. Prospective validation across heterogeneous clinical workflows is required. Fourth, the framework relies exclusively on H&E histology and does not incorporate clinical metadata (prior antibiotic use, treatment history) or molecular information. Integrating multimodal data may improve both predictive performance and biological interpretability. Fifth, the models showed moderate calibration error (ECE 0.1843 for CLA and 0.2126 for FQs), reflecting a tendency to overestimate probabilities in intermediate ranges. Although discrimination (AUC) and clinical net benefit remained strong, post hoc recalibration (eg, temperature scaling or isotonic regression) should be applied before the predicted probabilities are used for probability-threshold clinical decisions. Relatedly, resistance prevalence in the independent test set was moderate (CLA 35.8%, FQs 34.4%) with imbalanced class distributions; we therefore prioritized prevalence-independent discrimination and reported prevalence-adjusted predictive values, but validation in cohorts spanning a wider range of resistance prevalence is warranted. Conclusion We show that a weakly supervised deep learning framework applied to routine H&E-stained gastric biopsy WSIs can discriminate H. pylori antibiotic resistance phenotypes with strong overall performance. This approach is not intended to replace molecular susceptibility testing, but provides complementary, accessible decision support for resistance risk stratification—particularly in resource-limited settings—pending prospective validation. Funding Statement This study was supported by Futian Healthcare Research Projects (grant numbers FTWS035 and FTWS2025021). Data Sharing Statement The data that support the findings of this study are available from the corresponding authors (Chao Zeng and Yimin Guo) upon reasonable request. Owing to institutional and patient-privacy restrictions, a data sharing agreement must first be initiated between the interested parties and the clinical institution, following institution-specific guidelines. Ethics Approval and Consent to Participate The study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Boards of both participating institutions: The Eighth Affiliated Hospital of Sun Yat-sen University and The Second Affiliated Hospital of Guangzhou Medical University (protocol code 2025r259). Written informed consent was waived given the retrospective nature of the study. As patient consent to review medical records was not required by the ethics committees, all records and whole-slide images were fully de-identified prior to analysis, and patient data confidentiality was maintained throughout the study in accordance with the participating institutions’ data-protection policies. Author Contributions All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work. Disclosure The authors report no conflicts of interest in this work. References Associated Data Data Availability Statement The data that support the findings of this study are available from the corresponding authors (Chao Zeng and Yimin Guo) upon reasonable request. Owing to institutional and patient-privacy restrictions, a data sharing agreement must first be initiated between the interested parties and the clinical institution, following institution-specific guidelines.