Superpixel-graph self-supervised pretraining with joint-embedding and contrastive objectives for wound image segmentation

Annotated data remain a major problem and bottleneck for supervised learning approaches, especially in wound imaging.In this work, a self-supervised learning pretraining method was proposed that represents a wound image as a superpixel adjacency graph, generated using the simple linear iterative clustering algorithm and uses graph structure to define building blocks for calculating joint-embedding and contrastive learning objectives.Three objective variants were evaluated -contrastive, joint-embedding and their combination as a pretraining stage for further finetuning for the wound segmentation task.After self-supervised pretraining on 100 unlabelled images and supervised fine-tuning on only 30 labelled images using the contrastive objective, a Dice score of 0.266 was received on a 400-image test set, compared to 0.354 for a fully supervised U-Net trained on all 100 annotated images, thus having a gap of 0.088 in Dice score with more than three times less annotated data.The results suggest that superpixel-graph-based self-supervised learning is a promising pretraining strategy for wound analysis in settings with limited annotated data.