Bias reduction using combined stain normalization and augmentation for AI-based classification of histological images.

ABSTRACT

Artificial intelligence (AI)-assisted diagnosis is an ongoing revolution in pathology. However, a frequent drawback of AI models is their propension to make decisions based rather on bias in training dataset than on concrete biological features, thus weakening pathologists' trust in these tools. Technically, it is well known that microscopic images are altered by tissue processing and staining procedures, being one of the main sources of bias in machine learning for digital pathology. So as to deal with it, many teams have written about color normalization and augmentation methods. However, only a few of them have monitored their effects on bias reduction and model generalizability. In our study, two methods for stain augmentation (AugmentHE) and fast normalization (HEnorm) have been created and their effect on bias reduction has been monitored. Actually, they have also been compared to previously described strategies. To that end, a multicenter dataset created for breast cancer histological grading has been used. Thanks to it, classification models have been trained in a single center before assessing its performance in other centers images. This setting led to extensively monitor bias reduction while providing accurate insight of both augmentation and normalization methods. AugmentHE provided an 81% increase in color dispersion compared to geometric augmentations only. In addition, every classification model that involved AugmentHE presented a significant increase in the area under receiving operator characteristic curve (AUC) over the widely used RGB shift. More precisely, AugmentHE-based models showed at least 0.14 AUC increase over RGB shift-based models. Regarding normalization, HEnorm appeared to be up to 78x faster than conventional methods. It also provided satisfying results in terms of bias reduction. Altogether, our pipeline composed of AugmentHE and HEnorm improved AUC on biased data by up to 21.7% compared to usual augmentations. Conventional normalization methods coupled with AugmentHE yielded similar results while being much slower. In conclusion, we have validated an open-source tool that can be used in any deep learning-based digital pathology project on H&E whole slide images (WSI) that efficiently reduces stain-induced bias and later on might help increase pathologists' confidence when using AI-based products.