Predicting Lymph Node Metastasis From Primary Tumor Histology And Clinicopathologic Factors In Colorectal Cancer Using Deep Learning

This study proposes a method of controlling for known variables while selecting machine-learned features to develop a combined predictive model that maximizes generalizable performance and is inherently interpretable. The study utilized de-identified, digitized histopathology slides of primary colorectal samples and clinicopathologic metadata from colorectal cancer cases from two universities. The study found that this method of feature generation/selection does indeed provide a performance boost over known baseline variables that generalizes to an external dataset. Additionally, the study explored the ability of different deep learning pre-training regimens to generate relevant machine-learned features.