DU-Net based Unsupervised Contrastive Learning for Cancer Segmentation in Histology Images

TL;DR

This paper presents an unsupervised cancer segmentation method for histology images using a DU-Net architecture combined with contrastive learning, effective data augmentation, and CRF smoothing, outperforming some supervised methods.

Contribution

The introduction of a DU-Net based contrastive learning framework for unsupervised cancer segmentation in histology images, addressing the lack of annotated data.

Findings

Competitive segmentation performance surpassing some supervised networks

Effective use of data augmentation to enhance feature discriminability

Improved boundary detection with CRF smoothing

Abstract

In this paper, we introduce an unsupervised cancer segmentation framework for histology images. The framework involves an effective contrastive learning scheme for extracting distinctive visual representations for segmentation. The encoder is a Deep U-Net (DU-Net) structure that contains an extra fully convolution layer compared to the normal U-Net. A contrastive learning scheme is developed to solve the problem of lacking training sets with high-quality annotations on tumour boundaries. A specific set of data augmentation techniques are employed to improve the discriminability of the learned colour features from contrastive learning. Smoothing and noise elimination are conducted using convolutional Conditional Random Fields. The experiments demonstrate competitive performance in segmentation even better than some popular supervised networks.