Feature Selection-driven Bias Deduction in Histopathology Images: Tackling Site-Specific Influences

The emergence of bias in deep neural models represents a significant reliability concern, which may lead to overoptimistic results on seen data while compromising the model's ability to generalize effectively on unseen datasets. Recent studies conducted on The Cancer Genome Atlas (TCGA), which is a publicly available repository of histopathology images, reveal that the TCGA cancerous features extracted by deep neural networks surprisingly are able to discriminate slides based on their origin sites. This finding undoubtedly indicates the existence of site-specific patterns embedded in the extracted features learned by deep networks rather than focusing on histomorphologic patterns. Consequently, this biased behavior raises concerns about the reliability of these networks. This observation motivates us to conduct a series of experiments in which we present two distinct evolutionary feature selection strategies, each differentiated by its objective function evaluation. The primary goal is to select the features with a minimized foot-print of data source signatures, thereby ensuring a more accurate and site-independent cancer classification. We have conducted nine comprehensive independent experiments across nine cancer types, employing each evolutionary strategy. The comparison between results obtained through evolutionary strategies and the original feature sets highlights the substantial impact of feature selection methods on bias reduction while maintaining accuracy in cancer-type discrimination. Furthermore, the comparison of the two strategies with each other demonstrates the intricate nature of bias and its integration with cancerous features during the training process.