A Semi-Supervised Approach utilizing Deep Learning Pseudo-Label Refinement and Confidence Filtering for Enhanced Image Segmentation: An application towards Skin Lesion Analysis

Abstract

Accurate segmentation of skin lesions is an important step in diagnosing skin cancer. However, most segmentation methods based on deep learning rely heavily on large volumes of manually annotated images, which are both costly and difficult to obtain. This paper proposes a new framework for semi-supervised learning that effectively links an abundance of unlabelled thermoscopic images with a small, labelled dataset to improve the performance of segmentation. The framework integrates an architecture of Teacher-Student, where the Teacher network builds pseudo-labels for the unlabelled samples. The refinement of pseudo-labels is achieved through uncertainty-based filtering, which discards unreliable areas, and is further enhanced by Conditional Random Field (CRF) post-processing, which sharpens the lesion boundaries. The refined pseudo-labels are then utilised for training the student network, allowing for robust learning through both unlabelled and labelled data. Experiments on the HAM10000 and ISIC 2018 datasets indicate that the proposed model achieves an IoU of 0.64 and a Dice Coefficient of 0.74 with only 10 percent of labelled data. It was comparable to fully supervised training with 100 percent annotations. Qualitative outcomes suggest improved delineation of lesion boundary and decreased false segmentation. This research work demonstrates the potential of pseudo-label refinement as an effective strategy for addressing the scarcity of annotations in medical image segmentation.