Author: Muhammad Ramish Ashraf, Kerriann Casey, Suparna Dutt, Jie Fu, Edward Elliot Graves, Xuejun Gu, Hao Jiang, Brianna Caroline Lau, Billy W Loo, Weiguo Lu, Rakesh Manjappa, Stavros Melemenidis, Erinn Bruno Rankin, Lawrie Skinner, Luis Armando Soto, Murat Surucu, Vignesh Viswanathan, Zi Yang, Amy Shu-Jung Yu 👨🔬
Affiliation: Department of Radiation Oncology, University of Washington and Fred Hutchinson Cancer Center, Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, Department of Radiation Oncology, Stanford University School of Medicine, Department of Comparative Medicine, Stanford University School of Medicine, Department of Radiation Oncology, Stanford University Cancer Center 🌍
Purpose: The intestine is a classical preclinical model for studying radiation injury, and histological quantification of intestinal crypts is a key assay for assessing this response. However, substantial staining variations can hinder downstream analysis, particularly in deep learning (DL) applications. This study explores an unsupervised, task-specific DL approach for stain standardization and crypt auto-detection in hematoxylin-eosin (H&E)-stained histology images to enhance the evaluation of normal tissue FLASH irradiation response.
Methods: Our dataset includes 350 H&E-stained mouse intestinal histology images from a prior study on abdominal normal tissue response to FLASH irradiation. The images were categorized into two groups: 200 target images representing the desired stain style and 150 source images with suboptimal stain styles. The workflow involves stain normalization using an unsupervised CycleGAN, augmented with attention networks in each generator to produce attention maps that highlight discriminative regions and facilitate the translation of source style images into the target style. This is then coupled with a UNet-based network for crypt auto-detection on the translated images. Both networks are trained simultaneously to emphasize task-specific features such as crypts.
Results: Qualitatively, our approach demonstrates a significant reduction in staining variations and achieves standardized stain color distribution while maintaining structural integrity. Quantitatively, we assessed 9 source style images using our workflow, comparing crypt counts with those obtained from a crypt-counting model trained exclusively on target style images. While the original model detected only 28% of the crypts in source style images, our proposed workflow improved detection sensitivity to 81% after standardizing color variations.
Conclusion: Our unsupervised attention-guided DL approach can effectively standardize stain color distribution and reduce staining variations while preserving structural content. This method offers a promising tool for assessing normal tissue radiation responses in histology images.