Racially unbiased deep learning-based mammogram analyzer

Abstract

A racially unbiased mammogram analyzer includes an interface for receiving mammograms; a processor for extracting features of mammograms of general population; a processor for extracting features of mammograms of a specific race. In one embodiment, the general population mammogram features are represented by middle layers of a CNN and the race specific features are represented by the end layer of the CNN network. In one embodiment, the race specific layers of CNN change dynamically according to the race indication done explicitly. In one embodiment the race specific layers of CNN change dynamically according to the race indication given by race indication processor. In one embodiment, the race indications are computed by a network of parallel variational autoencoder networks. In one embodiment, the race indicator computes race specific information to the CNN and are provided by variational autoencoders.