|
Abstract:
|
Personalized medicine for cancer relies on accurate identification of subtype and mutations present in tumors. Currently, visual inspection of histopathology slides is one of the main methods used by pathologists to assess the stage, type and subtype of tumors. We present a series of Convolutional Neural Networks (CNNs) that have been successfully validated for lung cancer, endometrial cancer, and pan-cancer (over 28 cancers) subtype detection tasks. Our results indicate heldout validation Area Under ROC (AUC) of 0.97 for lung, an AUC of 0.969 for endometrial, and an average AUC of 0.989 for the pan-cancer subtype detection. Besides subtypes, genomics mutation identification is also an essential part of targeted therapy treatment. Our results indicate that CNN models are able to predict mutations in a tumor directly from H&E stained histopathology. In Lung cancer, mutations STK11, EGFR, FAT1, SETBP1, KRAS and TP53 can be predicted with AUCs from 0.733 to 0.856. Considering that the CNN models outperform both pathologists and traditional imaging methods, we will discuss model explanations to visualize the signatures of cellular phenotypes corresponding to our trained models.
|
