Self-Supervised Vision Transformers Learn Visual Concepts in Histopathology

TL;DR

This paper demonstrates that self-supervised Vision Transformers, especially DINO-based models, effectively learn interpretable morphological features in histopathology images, aiding tissue phenotyping in cancer research.

Contribution

The study introduces a comprehensive evaluation of self-supervised Vision Transformers in histopathology, highlighting DINO-based models' ability to learn diverse, interpretable features efficiently.

Findings

DINO-based Vision Transformers learn distinct morphological phenotypes.

Self-supervised models outperform traditional transfer learning methods.

Models provide interpretable attention maps for tissue analysis.

Abstract

Tissue phenotyping is a fundamental task in learning objective characterizations of histopathologic biomarkers within the tumor-immune microenvironment in cancer pathology. However, whole-slide imaging (WSI) is a complex computer vision in which: 1) WSIs have enormous image resolutions with precludes large-scale pixel-level efforts in data curation, and 2) diversity of morphological phenotypes results in inter- and intra-observer variability in tissue labeling. To address these limitations, current efforts have proposed using pretrained image encoders (transfer learning from ImageNet, self-supervised pretraining) in extracting morphological features from pathology, but have not been extensively validated. In this work, we conduct a search for good representations in pathology by training a variety of self-supervised models with validation on a variety of weakly-supervised and patch-level tasks. Our key finding is in discovering that Vision Transformers using DINO-based knowledge distillation are able to learn data-efficient and interpretable features in histology images wherein the different attention heads learn distinct morphological phenotypes. We make evaluation code and pretrained weights publicly-available at: https://github.com/Richarizardd/Self-Supervised-ViT-Path.