Equivariant Imaging Biomarkers for Robust Unsupervised Segmentation of Histopathology

TL;DR

This work tackles the challenge of rotational and reflection variability in histopathology by introducing SRENet, a rotation/reflection-equivariant CNN built on centrally symmetric convolutional kernels. It enables unsupervised learning of robust imaging biomarkers and unsupervised segmentation via K-means clustering on features from a rotation-equivariant backbone, demonstrated on 50 prostate TMAs with favorable intra- and inter-subject rotation robustness compared with ResNet and E2CNN. A downstream embedding-based evaluation shows stronger alignment with pathologist-defined labels, suggesting that the learned equivariant embeddings capture clinically relevant tissue structure. The approach holds promise for broader generalization across histopathology and for longitudinal analysis, while future work includes diverse pretraining and comparisons to capsule-based or transformer-based architectures.

Abstract

Histopathology evaluation of tissue specimens through microscopic examination is essential for accurate disease diagnosis and prognosis. However, traditional manual analysis by specially trained pathologists is time-consuming, labor-intensive, cost-inefficient, and prone to inter-rater variability, potentially affecting diagnostic consistency and accuracy. As digital pathology images continue to proliferate, there is a pressing need for automated analysis to address these challenges. Recent advancements in artificial intelligence-based tools such as machine learning (ML) models, have significantly enhanced the precision and efficiency of analyzing histopathological slides. However, despite their impressive performance, ML models are invariant only to translation, lacking invariance to rotation and reflection. This limitation restricts their ability to generalize effectively, particularly in histopathology, where images intrinsically lack meaningful orientation. In this study, we develop robust, equivariant histopathological biomarkers through a novel symmetric convolutional kernel via unsupervised segmentation. The approach is validated using prostate tissue micro-array (TMA) images from 50 patients in the Gleason 2019 Challenge public dataset. The biomarkers extracted through this approach demonstrate enhanced robustness and generalizability against rotation compared to models using standard convolution kernels, holding promise for enhancing the accuracy, consistency, and robustness of ML models in digital pathology. Ultimately, this work aims to improve diagnostic and prognostic capabilities of histopathology beyond prostate cancer through equivariant imaging.

Figures (6)

• Figure 1: Intra-Subject and Inter-Subject Analysis. We visualize an example image for intra-subject (A) and inter-subject (B) analyses using equivariant learning (SRENet and E2CNN) and standard convolution (ResNet). The TMA image undergoes 30-degree rotation increments (top row). For each rotation angle, the resulting segmentation after unsupervised K-means clustering was plotted and then unrotated back to the original input orientation for each model. Because the K-means cluster fitting was performed independently, the segmentation label colormap is not consistent across models.
• Figure 2: Comparison to Pathologist Segmentation. For 4 TMA image example subjects (column 1), we visualize the pathologist segmentations (column 2) in comparison with labeled segmentation maps from our equivariant SRENet model (column 3), rotation equivariant baseline E2CNN (column 4), and conventional non-equivariant baseline ResNet (column 5). Because the K-means cluster fitting was performed independently, the segmentation label colormap is not consistent across models or between models and pathologist segmentations.
• Figure A1: Intra-subject and Inter-subject ICC Analysis. Intra- and inter-subject class agreement analyses using equivariant learning (SRENet and E2CNN) and standard convolution (ResNet) evaluated with ICC. * indicates p$<$0.05.
• Figure A2: Quantitative Comparison to Pathologist Segmentation. We project ground-truth pathologist Gleason Grade labels onto a low-dimensional embedding of model imaging features for each patient in the test set and evaluat using Dice.
• Figure A3: Qualitative Comparison to Pathologist Segmentation. We project ground-truth pathologist Gleason Grade labels onto a low-dimensional embedding of model imaging features for each patient shown in Fig. \ref{['fig:path']}.