Bias Reduction in Representation of Histopathology Images Using Deep Feature Selection

Abstract

Appearing traces of bias in deep networks is a serious issue that can play a significant role in ethics and generalization. Recent studies report that the deep features extracted from the histopathology images of The Cancer Genome Atlas (TCGA), the largest publicly available archive of 11,000 patients covering 25 organs and 32 cancer subtypes, are surprisingly able to accurately classify the whole slide images (WSIs) based on their acquisition site. This is clear evidence that the utilized Deep Neural Networks (DNNs) unexpectedly detect the specific patterns of the source site rather than histomorphologic patterns, biased behaviour resulting in degraded generalization. This observation motivated us to propose a method to alleviate the destructive impact of hospital bias through a novel feature selection process. To this effect, we have proposed an evolutionary strategy select a small set of optimal features to not only accurately represent the histological patterns of tissue samples but also to eliminate the features leading to internal bias toward the institution. The performance of the proposed method has been assessed using the features extracted from TCGA images made available by NIH (National Institute of Health). The selected features extracted by a state-of-the-art network trained on TCGA images (i.e., the KimiaNet), considerably decreased the institutional bias. The proposed scheme is not limited to DNNs and/or specific types of biases; it can be employed to reduce various kinds of bias in a much more comprehensive range of data-driven feature extraction models. The conducted experiments, the external validation in specific, clearly demonstrate that the bias cannot be fully controlled just during training a model. Therefore, a feature selection for downstream tasks plays a crucial role in significantly reducing the bias.