by Christos Evangelou, MSc, PhD – Medical Writer and Editor

In a new study, researchers at Institut Universitaire du Cancer-Oncopole and Université Toulouse III-Paul Sabatier in Toulouse, France, developed methods for hematoxylin and eosin (H&E) color augmentation and image normalization to reduce bias caused by stain color variations in histopathology images used for machine learning.¹

The team evaluated their methods on a breast cancer dataset with images from three clinical trials that exhibited stain color bias and found that the combination of color augmentation and image normalization reduced stain bias during AI training.

The authors conclude that the proposed methods for stain augmentation and normalization could help mitigate a primary source of bias in using machine learning on histopathology images from different sources.

“By providing data scientists with two effective bias reduction tools, this will make the development of AI tools in histopathology easier, with tangible clinical impact,” said Camille Franchet, MD, co-first and the corresponding author of the study.

The report was published in Computers in Biology and Medicine.

Rationale: Reducing Bias in AI-Assisted Histopathology Image Analysis

Histopathology relies heavily on the ability to distinguish different structures and cellular components. This process is facilitated by staining techniques, most commonly using H&E. However, the specific staining protocols and dye compositions can vary significantly between laboratories, resulting in substantial color differences in the stained tissue samples.¹

This variability in tissue staining techniques across laboratories poses a significant challenge for AI models trained on datasets from different sources. Because of stain variability in datasets used to train AI models, algorithms may fail to generalize and accurately analyze images with different staining characteristics. This limitation has hindered the widespread adoption of AI in digital pathology, as pathologists are cautious about relying on models that may be biased or inconsistent.¹

“Given the often limited size of datasets and the challenge to work with widely multicentric data, it appears that the performance of deep learning-based AI models in histopathology could be biased due to slides staining variations, resulting in models with poor generalizability,” noted Dr. Franchet. “All this has been observed during the development of a classification tool concerning histological images, which led us to increase our efforts on mitigating these biases.”

A Two-Pronged Approach: Augmentation and Normalization

To reduce bias and increase the reliability of AI-based analysis of histopathology images, the team has developed two innovative methods: AugmentHE for stain color augmentation and HEnorm for H&E image normalization.¹

AugmentHE is a novel approach to data augmentation, a technique commonly used in machine learning to expand the diversity of training data. By separately modifying the hematoxylin and eosin color channels, AugmentHE generates realistic augmented histology images that mimic the natural variability in staining protocols.

To facilitate fast and efficient H&E image normalization, the team developed HEnorm, a deep learning-based normalization method. Unlike conventional normalization techniques that can be computationally expensive, HEnorm utilizes the hematoxylin channel to reconstruct the normalized H&E image.

Commenting on the novelty of this approach, co-first author Robin Schwob, MSc, said: “We leveraged our precise knowledge of slide staining process to develop both normalization and color augmentation methods, which take advantage of all aspects of hematoxylin and eosin channels for deconvolution to accurately represent staining variations across laboratories.”

Method Validation

To evaluate the effectiveness of their methods, the researchers employed a breast cancer dataset comprising images from three different clinical trials, each stained at a distinct central laboratory. This setup provided an ideal testbed for assessing the ability of AugmentHE and HEnorm to mitigate bias due to stain variability.

The researchers trained classification models on images from a single clinical trial and then assessed their performance on the remaining datasets. By applying various combinations of augmentation and normalization techniques, including their own methods, they could monitor the impact on bias reduction and model generalizability.

Color augmentation using AugmentHE increased color dispersion by 81% compared to traditional geometric augmentations alone. In addition, H&E image normalization using HEnorm achieved up to 78 times faster processing while preserving image structure.¹

When used together, AugmentHE and HEnorm improved the area under the receiver operating characteristic curve (AUC) for a grade I vs III classification task by up to 21.7% on the biased test sets, compared to using only geometric augmentations. Moreover, the optimal approach was to first apply HEnorm for normalization, followed by AugmentHE for augmentation, yielding the best results for reducing stain bias during training.

“We extensively monitored bias reduction while providing accurate insight into both AugmentHE and HEnorm methods. Our pipeline improved AUC on biased data by nearly 22% compared to usual augmentations,” said Dr. Franchet. “This validated open-source tool can be used in any deep learning-based digital pathology project on H&E whole slide images in order to efficiently reduce stain-induced bias.”

Looking Ahead

“Our data normalization and augmentation tools might benefit from comparison with more computationally intensive methods such as generative adversarial networks,” Schwob acknowledged.

He added that although predominant, slide staining is not the sole source of bias in microscopic image datasets. “Slice thickness or variations in the pre-analytical stage might also introduce systematic biases,” Schwob explained.

The authors also noted that as the field of digital pathology continues to evolve, the need for unbiased and generalizable AI models will only become more pressing. This study represents a significant step forward, but further research and collaboration between pathologists, computer scientists, and industry partners will be essential to fully realize the potential of AI in histopathology image analysis.

The study was funded Bpifrance (APRIORICS project), Thales Services Numérique, Fondation pour la Recherche Médicale, France, Agence Nationale de la Recherche, France, Institut Carnot CALYM (DIAL project) and Laboratoire d’Excellence Toulouse Cancer (TouCan). References

1. Franchet C, Schwob R, Bataillon G, et al. Bias reduction using combined stain normalization and augmentation for AI-based classification of histological images. Comput Biol Med. 2024;171:108130. doi:10.1016/j.compbiomed.2024.108130