Skip to main navigation menu Skip to main content Skip to site footer

Abstracts & Posters

Vol. 1 No. S1 (2025): Special Issue: 2025 Geisel Research Poster Night

Interpretable Classification of Renal and Colonic Neoplasia Using Prototypical Explainability

Submitted
19 December 2025
Published
27-12-2025

Abstract

Interpretable artificial intelligence tools are increasingly needed in pathology to overcome the opacity of black-box models and improve clinician trust. This study applied a prototypical part network—a deep learning architecture that compares regions of an input image to learned prototypes—to classify renal cell carcinoma (five subtypes) and colorectal lesions (hyperplasia vs. sessile serrated adenoma). Trained on DHMC datasets of whole-slide images and biopsy patches, the model achieved accuracies of 92.3% (renal) and 85.5% (colonic). The prototype-based approach enabled visualization of class-specific features, assisting in understanding both correct and incorrect predictions. Subpatch-level activation maps highlighted regions of suspected malignancy, aligning closely with pathologist annotations. This transparent decision-making framework may enhance diagnostic confidence and supports future clinical deployment of AI pathology tools.

References

  1. References are available on the poster PDF.