Efficient Self-Supervised Barlow Twins from Limited Tissue Slide Cohorts for Colonic Pathology Diagnostics

TL;DR

An optimized Barlow Twins framework for colorectal polyps screening is proposed and it is shown that the SSL representations are more meaningful and qualitative than the supervised ones and that Barlow Twins benefits from the Swin Transformer when applied to pathology data.

Abstract

Colorectal cancer (CRC) is one of the few cancers that have an established dysplasia-carcinoma sequence that benefits from screening. Everyone over 50 years of age in Canada is eligible for CRC screening. About 20\% of those people will undergo a biopsy for a pre-neoplastic polyp and, in many cases, multiple polyps. As such, these polyp biopsies make up the bulk of a pathologist's workload. Developing an efficient computational model to help screen these polyp biopsies can improve the pathologist's workflow and help guide their attention to critical areas on the slide. DL models face significant challenges in computational pathology (CPath) because of the gigapixel image size of whole-slide images and the scarcity of detailed annotated datasets. It is, therefore, crucial to leverage self-supervised learning (SSL) methods to alleviate the burden and cost of data annotation. However, current research lacks methods to apply SSL frameworks to analyze pathology data effectively. This paper aims to propose an optimized Barlow Twins framework for colorectal polyps screening. We adapt its hyperparameters, augmentation strategy and encoder to the specificity of the pathology data to enhance performance. Additionally, we investigate the best Field of View (FoV) for colorectal polyps screening and propose a new benchmark dataset for CRC screening, made of four types of colorectal polyps and normal tissue, by performing downstream tasking on MHIST and NCT-CRC-7K datasets. Furthermore, we show that the SSL representations are more meaningful and qualitative than the supervised ones and that Barlow Twins benefits from the Swin Transformer when applied to pathology data. Codes are avaialble from https://github.com/AtlasAnalyticsLab/PathBT.

Figures (26)

• Figure 1: Five slides from the five classes with their annotations (in black and blue) and patches from the ROI for the four polyps and from tissue regions for the Normal WSI. We observe that the annotations are not complete, as only one layer of the sample has been annotated (HP5, SSL6, TA3 and TVA2). The normal slide N9 does not present with any annotations.
• Figure 2: Impact of the different components of the augmentation strategy on the performance. We observe that the baseline, proposed in the original work BT does not perform well on our benchmark dataset.
• Figure 3: Two sets of new transformations are applied to the input of Barlow Twins to create distorted views. Examples of each transformation applied to different types of tissues are given as examples.
• Figure 4: The WSI are tiled into patches from ROI annotations or non-annotated regions. All patches are forwarded into the Barlow Twins framework. The input patches undergo two sets of distortions. The distorted patches are forwarded through an encoder, and their representations are projected into the loss space using a projector. The Barlow Twins loss $L_{BT}$ aims at making the cross-correlation matrix of these two embeddings close to the identity matrix. The learned representations are then evaluated on the patch level by training a Fully Connected Layer on top of the linear encoder and on the slide level using CLAM
• Figure 5: Polar graphs of the Accuracy and AUC of the patch classification for the different models and datasets. The proposed models, in blue and purple, perform relatively well and outperform other models with the same encoder for three out of the four datasets.