AI-Driven Digital Pathology: Advancing Precision Oncology
Artificial Intelligence (AI) on pathology specimens has the potential to accurately identify cancer site, classify cancer subtype and predict treatment effect. Recent studies suggest that AI methods can also make molecular predictions by learning genotype-phenotype relationships from pathology slides. Since molecular profiling is crucial for treatment and prognostic stratification, using AI-based methods on pathology images could potentially permit early understanding of individual tumor profiles prior to performing genomic tests. As such, AI-driven digital pathology has the potential to provide a cost-effective and time-efficient way to allocate individualized genomic tests and treatments.
In the first part of this seminar, I will discuss current trends in AI for digital pathology and practical applications. In the second part, I will present our interpretable AI framework for genomic alteration prediction using high-resolution hematoxylin and eosin (H&E) whole slide images. In our work, we first used pretrained pathology foundation models to extract features from the H&E images and then trained a multiple instances learning (MIL) model to predict specified genomic alternations as well as microsatellite instability (MSI) and tumor mutation burden (TMB). Additionally, we assessed the performance of using different foundation models as the feature extractor and compared various approaches for downstream genomic alteration prediction. The models are trained and cross-validated using biopsy samples from 203 newly diagnosed prostate cancer patients from University of Washington OncoPlex clinical biopsy cohort. We find heterogenous performances across different gene alterations but whether this is due to data, algorithm, or biology remains to be determined. Further research is needed to discover reliable associations between learned morphological patterns and molecular alterations.