Pancreatic ductal adenocarcinoma (PDAC) is not a single disease but encompasses at least two molecularly distinct subtypes: the classical subtype, which tends to express genes related to epithelial identity and may be more responsive to certain chemotherapy regimens, and the basal-like (squamous) subtype, which shows features of epithelial-to-mesenchymal transition, is more aggressive, and has a worse prognosis.
The most widely used molecular classification tool is PurIST (Purity Independent Subtyping of Tumors), an RNA-expression-based algorithm that assigns PDAC samples to classical or basal subtypes based on the relative expression of a defined gene panel. However, PurIST requires fresh frozen tissue and RNA sequencing, resources not available in all clinical settings.
If the same subtype information could be extracted directly from routine whole slide pathology images (WSI) - the same hematoxylin-and-eosin slides already produced for diagnosis - subtyping could be done without extra molecular testing, dramatically expanding access to precision treatment guidance worldwide.
The study developed a two-stage classifier pipeline. In the first stage, a patch-level convolutional neural network was trained to identify which regions of a whole slide image contain tumor tissue versus stroma, necrosis, or normal pancreas - a process called tissue segmentation. This ensures that the subtyping model focuses on informative tumor patches rather than being confused by non-tumor content.
In the second stage, a separate deep learning model - the Vanilla model - was trained on patches from tumor regions to predict PurIST subtype labels. The training dataset came from TCGA-PAAD (The Cancer Genome Atlas Pancreatic Adenocarcinoma), the largest publicly available collection of PDAC specimens with matched histopathology images and molecular subtype calls.
To test generalization, the trained models were applied to an external validation cohort of locally collected biopsy specimens - cases the model had never seen during training. Performance was measured by comparing model-assigned subtypes to PurIST calls from RNA sequencing of the same specimens, providing an objective benchmark against the molecular gold standard.
On the TCGA training and internal test set, the Vanilla deep learning model achieved 96.19% accuracy in assigning classical versus basal-like subtype labels, demonstrating that morphological features in H&E images carry substantial molecular subtype information detectable by AI.
On the external biopsy validation cohort - a more stringent test that reflects real-world deployment conditions - accuracy remained high at 83.03%. The drop from 96% to 83% between internal and external validation is expected given differences in tissue processing, staining protocols, and scanner hardware across institutions, and the external accuracy is still well above chance.
The finding that a model trained on one institutional dataset retains most of its accuracy when applied to a different institution's specimens is a critical indicator of model robustness and potential clinical deployability. Many AI pathology models fail catastrophically on external data due to overfitting; this model's retained accuracy suggests it has learned biologically meaningful features rather than institution-specific artifacts.
Although deep neural networks are often considered black boxes, visualization techniques such as attention maps and gradient-weighted class activation mapping (Grad-CAM) can highlight which image regions most influenced the model's predictions. In PDAC subtyping, the model tended to focus on nuclear morphology, gland architecture, and the density and arrangement of tumor cells - features that experienced pathologists also use to subjectively distinguish well-differentiated from poorly differentiated tumors.
The overlap between AI-identified features and established pathological criteria provides reassurance that the model is capturing genuine biological differences rather than technical noise. It also opens the door to discovering new morphological descriptors of subtype that human observers might not have previously articulated.
The classical subtype was associated with more organized glandular structures and mucin production, while the basal-like subtype showed more solid tumor sheets, nuclear pleomorphism, and reduced glandular differentiation - patterns consistent with the known molecular characteristics of each subtype and with their respective clinical behaviors.
Molecular RNA-based subtyping currently requires specialized laboratory infrastructure - fresh frozen tissue storage, RNA extraction facilities, and bioinformatics pipelines - that is available at major academic cancer centers but not at community hospitals or institutions in low-resource settings. Deep learning-based subtyping from standard H&E slides requires only a digital slide scanner, eliminating this infrastructure barrier.
Clinical implementation would allow oncologists to select subtype-matched chemotherapy regimens at the time of initial diagnosis. Early evidence suggests that basal-like PDAC responds differently to FOLFIRINOX versus gemcitabine plus nab-paclitaxel, and routine subtype identification could guide these first-line treatment choices in a meaningful way for patient outcomes.
As larger molecularly annotated PDAC cohorts become available for training, accuracy is likely to improve further. Future models might also predict additional molecular features beyond two-class subtyping - such as specific pathway activations, mutational signatures, or immunotherapy response markers - expanding the information that can be extracted from a single diagnostic slide at no extra procedural cost.