A self-supervised representation learning model for biopsy images of the human duodenum using super-pixel inpainting. In this method, we design a novel composite pretext task for targeted inpainting and elastic deformations. A fully supervised model is initially trained on a small number of labeled data. This model is used to generate psuedo-labels for important anatomical regions in unlabeled biopsy images. Some of these regions are randomly masked and the entire image is deformed using elastic deformation as mentioned in this paper. An encoder-decoder pair is trained to reconstruct the masked, deformed image using the SSIM loss. The model thus learns representation for shapes and H&E color stain distributions of different important tissues like Epithelium, Crypts and Villi. This model is then fine-tuned on a small labeled subset of images for semantic segmentation.
Visualization of representation learning:
