Comparison of Digital Histology AI Models with Low-Dimensional Genomic and Clinical Models in Survival Modeling for Prostate Cancer

Abstract

Recent advances in AI have transformed digital pathology cancer diagnosis tasks, such as weakly supervised prediction of subtype or disease trajectory from unannotated whole slide images (WSIs). Previous works evaluate overall survival in certain cancers instead of progression end-points, which can have more clinical relevance in prostate cancer. In addition, most of these studies only train uni-modal WSI models, omitting clinical-pathological and gene expression data from the analysis. Moreover, batch effects in WSI are largely ignored in prior works. Our work aims to integrate these evaluations in the context of predicting progression-free survival using the Prostate Adeno-carcinoma dataset (PRAD) from The Cancer Genome Atlas (TCGA). Our results show that Digital Pathology AI (DPAI) models perform similarly to multivariate Cox models of gene expression or clinical features such as Gleason Group. The predicted risk scores of the genomic unimodal model are weakly correlated with those of the DPAI and clinical models. While a combined clinical genomic model shows enhanced performance, multimodal models with DPAI, genomic, and clinical features show similar performance to unimodal DPAI models in this dataset. Among the four DPAI architectures we explored, comparable performance was observed when local patch aggregation is part of the architecture or when a DPAI foundation model is employed as the feature extractor. We also demonstrate that ignoring batch effects can lead to worse performance. Our work highlights a critical need for larger public WSI datasets to properly evaluate DPAI prognostic model architectures.

Ähnliche Arbeiten