Artificial intelligence–driven quantification of tumor-stroma ratio reveals prognostic significance of tumor burden and ductal morphology in pancreatic ductal adenocarcinoma

Technical Summary: AI-Driven Quantification of Tumor-Stroma Ratio in Pancreatic Ductal Adenocarcinoma

Problem Statement
Pancreatic ductal adenocarcinoma (PDAC) is an aggressive malignancy with poor prognosis, largely driven by early metastasis and therapy resistance. While the tumor-stroma ratio (TSR) and specific histological growth patterns (such as ductal morphology) are recognized prognostic factors, their clinical utility is hindered by the subjectivity and limited scalability of conventional visual assessment on hematoxylin–eosin (H&E) stained slides. Existing molecular classification systems lack sufficient robustness for routine clinical decision-making. Consequently, there is a critical need for objective, quantitative, and reproducible methods to evaluate TSR and ductal morphology to refine risk stratification in PDAC.

Methodology
The study developed a high-precision digital pathology pipeline centered on a ResNet-50 deep learning architecture to automate the quantification of histological components in PDAC.

• Data Cohort: The study utilized 740 H&E-stained whole-slide images (WSIs) from 320 PDAC patients treated at Huashan Hospital, Fudan University, between 2019 and 2022. External validation was performed using cohorts from Zhongshan Hospital, the First Affiliated Hospital of Naval Medical University, and The Cancer Genome Atlas (TCGA).
• Deep Learning Model: A ResNet-50 model, employing a Bottleneck Block structure and identity mappings, was trained to classify nine distinct histological components: tumor epithelium, stroma (fibrous tissue), pancreatitis, necrosis, normal acini, islets, lymphoid tissue, intestinal mucosa, and adipose tissue.
• Training and Validation: The internal cohort was split 7:3 into training (n=229) and validation (n=91) sets. The model was trained using cross-entropy loss with stochastic gradient descent (SGD), incorporating data augmentation (rotation, flipping, color jittering). Performance was evaluated using confusion matrices, accuracy, precision, recall, and ROC curves.
• Quantitative Metrics: The system calculated the Tumor-Stroma Ratio (TSR), tumor burden (relative proportion of tumor patches), and stromal content. Additionally, ductal morphology was quantified using QuPath software to classify tumors as Large-Duct Predominant (LDP, ducts >0.5 mm) or Small-Duct Predominant (SDP).
• Prognostic Modeling: Five machine learning algorithms (Random Forest, XGBoost, k-nearest neighbors, CatBoost, and logistic regression) were employed to predict Overall Survival (OS) and Progression-Free Survival (PFS). Hyperparameter optimization was conducted using Optuna. Model interpretability was achieved via SHAP (SHapley Additive exPlanations) analysis and feature interaction studies.

Key Results

• Model Performance: The ResNet-50 classifier demonstrated high accuracy, achieving an overall multiclass ROC of 0.94 in external validation cohorts and 0.97 in the TCGA cohort. Specific recognition accuracy for tumor epithelium reached 0.98, and for stromal tissue, it reached 0.90 (external) and 1.00 (TCGA). The primary source of misclassification was the similarity between pancreatitis and tumor epithelium.
• TSR and Prognosis: A high TSR was significantly associated with shorter Progression-Free Survival (PFS; p = 0.0002) and Overall Survival (OS; p = 0.026). High TSR also correlated with higher histological grade, vascular tumor thrombi, and elevated Ki-67 indices.
• Dissecting TSR Drivers: Crucially, the study found that the adverse prognostic impact of a high TSR was primarily driven by increased tumor burden, not stromal abundance alone. Patients with high tumor burden (highest quartile) had significantly shorter PFS, whereas stromal content alone showed no significant association with survival outcomes.
• Ductal Morphology: Tumors exhibiting a Large-Duct Predominant (LDP) morphology were associated with lower TSR, less aggressive histopathological features, and significantly improved OS (p = 0.015) and PFS (p = 0.026) compared to Small-Duct Predominant (SDP) tumors.
• Machine Learning Insights: Among the tested algorithms, the optimized Random Forest model performed best. SHAP analysis identified N stage, histological grade, and pT stage as the strongest predictors for OS. For PFS, tumor burden, necrotic proportion, and TSR were the primary drivers of risk.

Significance and Claims
The paper claims that the proposed ResNet-50-based digital pathology framework enables precise, automated, and reproducible quantification of PDAC histological components, overcoming the interobserver variability inherent in manual assessment.

The study establishes that:

1. High TSR is a robust prognostic marker for unfavorable outcomes in PDAC, but its clinical utility is clarified by the finding that it serves as a proxy for tumor burden rather than stromal density.
2. Ductal morphology (LDP vs. SDP) provides complementary prognostic value, with LDP delineating a subset of PDAC with less aggressive biological behavior.
3. Integration of AI and Histopathology offers a scalable approach to risk stratification that complements existing pathological assessments.

The authors acknowledge limitations, including the retrospective nature of the study, the reliance on histopathological images without molecular integration, and the inherent complexity of deep learning models. They conclude that while the system provides a mechanistic framework for risk stratification, future work requires prospective validation in larger, multi-institutional cohorts to fully integrate these findings into clinical decision-making.