Multistage Image Reconstruction and Attention-Based Semi-Supervised Learning for Medical Image Segmentation

Abstract

Automated segmentation of medical images is critical in detecting and diagnosing various conditions. In recent years, supervised deep learning (DL) techniques have been widely researched. However, their application is often limited by the availability of annotated data in the medical domain. To address this, recent studies have explored semi-supervised techniques, though very few of these works focus on skin-lesion segmentation. In addition, they struggle to effectively capture contextual features to delineate the region of interest from the surrounding tissues in the image, which is crucial for accurate segmentation. In this article, a semi-supervised approach for medical image segmentation called MIRA (Medical Image Reconstruction and Analysis) is proposed, which uses adaptive-attention U-Net (AA-U-Net) trained on pseudo-labels generated with a lightweight feature-consistent encoder-decoder network (FCED-Net) to address these challenges. A case study focusing on the precise segmentation of malignant skin lesions is considered for our experiments, as the scarcity of extensive annotated dermatology data limits the effectiveness of traditional DL models. The proposed pipeline is validated and tested using two standard datasets, ISIC2016 and PH². With only 50% annotated samples, the proposed approach demonstrated promising performance with DSC, IoU, and accuracy of 0.96, 0.92, and 0.85 on ISIC2016 and 0.93, 0.88, and 0.93 on cross-data testing with PH ${}^2$ dataset. When benchmarked against leading edge models trained on 100% labeled data, MIRA achieved promising results and even outperformed in some cases. These findings show that it can significantly reduce manual annotation requirements while achieving segmentation performance comparable to models trained on fully annotated skin lesion data.