Bridging traditional and deep learning methods in H&E histological image normalization: a comprehensive review and introducing a novel framework for comparative analyses

Abstract

Background

Histology images are a cornerstone of pathology, which allow automated analysis for disease diagnosis. However, variations in staining and image acquisition processes significantly affect the performance of these algorithms. Histology image normalization is method to achieve uniformity in image color distributions, which will enhance the accuracy and consistency of automated analysis.

Aim of review

This review was conducted with the aim of assessing normalization methods and comparing them in an empirical manner to help researchers choose the most appropriate method for their study. It also aims to assist academics and professionals involved in automated image analysis and digital pathology.

Key scientific concepts of review

This review categorizes normalization techniques into four groups: deep learning-based approaches (e.g., GANs, autoencoders, diffusion models), traditional methods (e.g., deconvolution, histogram matching), hybrid models, and a novel signal processing-based method. It also introduces a new deep learning framework for evaluating normalization strategies and experimentally compares eight state-of-the-art methods on histopathology images. The results highlight the strengths and limitations of each approach, helping researchers and professionals choose suitable methods for their needs. In addition, the review emphasizes the impact of color variation on the accuracy of computer-aided diagnosis (CAD) systems and the importance of preserving biological information during normalization. Finally, it outlines directions for future research, including integrating normalization with data augmentation and exploring information preservation beyond cancer subtype.