Pancreatic ductal adenocarcinoma (PDAC) is one of the deadliest cancers worldwide, with a 5-year survival rate below 5% and an average survival of only 4.6 months after diagnosis. Europe alone sees around 150,000 new cases and 95,000 deaths per year, making better tools for predicting patient outcomes urgently needed.
Today, most treatment decisions rely on the TNM staging system, which grades tumors by size, lymph node involvement, and distant spread. However, patients with the same TNM stage often have very different outcomes, highlighting the system's limitations in guiding truly personalized care.
Researchers have explored the tumor-stroma ratio (TSR) - the proportion of tumor cells versus surrounding fibrous tissue called stroma - as a potential biomarker. A higher stromal content has been linked to worse prognosis in several cancers, but manual assessment of TSR is subjective and prone to variation between reviewers. This study aimed to automate TSR measurement using artificial intelligence.
The research team designed a multi-step pipeline that works on whole-slide images (WSIs) - digital scans of tissue sections stained with Hematoxylin and Eosin (H&E), the standard method for examining cancer tissue under a microscope. The pipeline automatically identifies tumor cells, maps the tumor boundary, and measures the stroma within it.
Four independent datasets were used: 16 patients from Radboud University Medical Center with paired H&E and immunohistochemistry (IHC) staining, 162 multicenter PDAC patients from a Dutch national pathology database, 29 patients from Radboud for internal validation, and 161 patients from TCGA (The Cancer Genome Atlas) for external validation and survival analysis.
The pipeline used U-Net neural networks with an EfficientNet-b0 encoder pretrained on ImageNet. Step one trained on IHC-stained slides to teach the network to recognize epithelial cells. Step two trained on the larger dataset to distinguish tumor epithelium from normal epithelium. A convex hull algorithm then automatically drew the tumor boundary, and a multi-tissue segmentation tool measured the stroma within that boundary to compute TSR.
Performance was measured using the Dice Similarity Coefficient (DSC), which compares the model's segmentation to expert pathologist annotations. Scores range from 0 to 1, with higher scores indicating better agreement. For survival prediction, a logistic regression model was trained with five-fold cross-validation to predict death within 6, 12, or 18 months.
The epithelium segmentation model achieved median Dice scores of 0.749 on training data and 0.717 on the external TCGA dataset, demonstrating consistent performance across different data sources. The model correctly identified tumor epithelial cells while excluding non-cancerous tissue like duodenal epithelium.
For tumor bulk segmentation, the convex hull approach achieved Dice scores of 0.700, 0.760, and 0.863 on the three test datasets respectively. The best performance on the TCGA dataset was partly due to larger tumors covering more of the tissue area in that cohort, making boundaries easier to define.
When TSR was used to predict 6-month survival, the logistic regression model achieved an AUC of 0.61 plus or minus 0.12, indicating moderate but meaningful predictive ability. Performance declined for longer survival windows (12 and 18 months), suggesting TSR is most informative for near-term outcomes. The mean automatically computed TSR was 0.731, with a median of 0.852, showing that most PDAC tumors in the cohort had very high stromal content.
The survival prediction AUC of 0.61 is modest, reflecting the well-known difficulty of prognosticating in pancreatic cancer. The authors note this is a preliminary study and that TSR alone is not sufficient as a standalone predictor - it works best when combined with clinical variables like age, tumor origin, and prior malignancies.
A key limitation is that the model was evaluated mainly on the TCGA dataset, which has its own biases in patient selection and tissue preparation. The relatively small external validation set (161 patients, 187 slides) limits the statistical confidence of the survival analysis results.
The pipeline also had difficulty distinguishing cancerous pancreatic epithelium from duodenal epithelium because both have similar glandular structures. Future improvements could include better training data, attention mechanisms to focus on the most relevant tissue regions, and validation on larger prospective cohorts with standardized treatment protocols.
This study demonstrates that fully automated TSR quantification from routine H&E tissue slides is feasible using a deep learning pipeline. By removing the need for manual pathologist scoring, the approach eliminates inter-observer variability and could enable consistent TSR measurement at scale across large patient cohorts.
The results support TSR as a computationally extracted biomarker with potential prognostic value in pancreatic cancer - a disease where reliable biomarkers have been notoriously hard to identify. The work follows successful automated TSR methods in colorectal cancer and extends them to the more complex PDAC microenvironment.
The research team made their code publicly available on GitHub, supporting open science and enabling other groups to build on or validate the approach. The next step would be larger validation studies and eventually integration with clinical workflows to help oncologists make better-informed treatment decisions for pancreatic cancer patients.