Identify morphological features that correspond to specific genetic alterations in lung cancer histopathology images

Characterize the precise morphological features in whole-slide hematoxylin and eosin histopathology images of lung adenocarcinoma that correspond to specific genetic alterations in driver genes TP53, EGFR, KRAS, and ALK, in order to improve the interpretability and clinical trust of deep learning-based mutation prediction models trained on the PathGene dataset.

Background

The paper introduces PathGene, a multicenter dataset linking lung cancer whole-slide histopathology images to next-generation sequencing reports, and benchmarks multiple instance learning models for predicting driver gene mutations, mutation subtypes, exons, and TMB status. While the models achieve strong predictive performance, the authors highlight interpretability challenges.

Specifically, the authors note that it is frequently unclear which morphological features in the images correspond to particular genetic alterations. This uncertainty limits clinical trust and hinders the adoption of AI models in precision oncology. Establishing a clear mapping from visually discernible histological patterns to specific mutations would advance explainability and facilitate clinical integration.