Bias in Histopathology Datasets: A Comprehensive Investigation on Possible Factors

Abstract

Deep Neural Networks (DNNs) have demonstrated remarkable capabilities in medical applications, including digital pathology, where they excel at analyzing complex patterns in medical images to assist in accurate disease diagnosis and prognosis. However, concerns have arisen about potential biases in The Cancer Genome Atlas (TCGA) dataset, a comprehensive repository for digitalized histopathology and a validation source for deep models, suggesting that over-optimistic results of model performance may be due to reliance on biased features rather than histological characteristics. Surprisingly, recent studies have confirmed the existence of site-specific bias in the embedded features extracted for cancer-type discrimination, leading to high accuracy in acquisition site classification. This biased behavior motivated us to conduct an in-depth analysis to investigate potential causes behind this unexpected biased ability toward site-specific pattern recognition. The analysis was conducted on two cutting-edge DNN models: KimiaNet, a state-of-the-art DNN trained on TCGA images, and the self-trained EfficientNet. Considering the classification accuracy of acquisition sites as a measure of bias, a series of experiments has been designed to examine the impact of various factors, including patching, data imbalance, and staining.