Integrating unsupervised visual representation learning on whole-slide images of bile duct strictures for enhanced diagnostics and prognostics.

Abstract

481 Background: Biliary obstructive disorders represent a major global health burden. Endoscopic retrograde cholangiopancreatography (ERCP) remains the standard diagnostic procedure. While ERCP provides morphologic imaging to differentiate benign from malignant strictures, definitive diagnosis still relies on tissue sampling. Conventional pathology of bile duct biopsies has limited sensitivity, often causing diagnostic uncertainty. These limitations highlight the need for AI-driven image analysis to enhance histologic evaluation and improve patient outcomes. Methods: We applied a self-supervised learning (SSL) framework combined with Uniform Manifold Approximation and Projection (UMAP), a non-linear dimensionality reduction technique preserving local and global structure, to classify bile duct histopathology images. The dataset included 78 whole-slide images (62 malignant, 16 benign), generating 94,744 patches. Patches were split at the patient level into training (70%) and validation (30%) sets, and the model was trained for 100 epochs. High-dimensional patch features were embedded into two-dimensional UMAP space for visualization. Unsupervised clustering revealed distinct, clinically meaningful tissue groupings, validated by expert pathologists. Results: SSL effectively distinguished histologically distinct tissue regions in bile duct WSIs. UMAP embeddings demonstrated strong neighborhood preservation (trustworthiness = 0.94, continuity = 0.89). Cluster quality metrics were robust (silhouette score = 0.63, Davies–Bouldin index = 0.47). Visualization revealed 24 clusters including invasive carcinoma, high- and low-grade biliary dysplasia, and associated stromal patterns. The model captured stromal heterogeneity, such as fibroblast morphology, collagen architecture, and immune infiltration, while distinguishing tumor subtypes by morphology. Non-informative background patches were excluded, and clinically relevant regions prioritized. Pathologist review confirmed clusters were interpretable and aligned with histological subtypes relevant for diagnosis and disease characterization. Conclusions: Beyond accelerating diagnostics, this method provides a framework for biomarker discovery. By capturing morphologic heterogeneity without predefined labels, unsupervised clustering can reveal tissue signatures potentially associated with relapse, survival, or therapeutic response. Integrating these clusters with clinical and molecular outcomes may enable new prognostic and predictive biomarkers, advancing precision medicine in cholangiocarcinoma.