Portable Skin Lesion Segmentation System with Accurate Lesion Localization Based on Weakly Supervised Learning

Abstract

The detection of skin lesions involves a resource-intensive and time-consuming process, necessitating specialized equipment and the expertise of dermatologists within medical facilities. Lesion segmentation, as a critical aspect of skin disorder assessment, has garnered substantial attention in recent research pursuits. In response, we developed a portable automatic dermatology detector and proposed a dual-CAM weakly supervised bootstrapping model for skin lesion detection. The hardware system in our device utilizes a modular and miniaturized design, including an embedded board, dermatoscope, and display, making it highly portable and easy to use in various settings. Our software solution uses a convolutional neural network (CNN) with a dual-class activation map (CAM) weakly supervised bootstrapping model for skin lesion detection. The model boasts two key characteristics: the integration of segmentation and classification networks, and the utilization of a dual CAM structure for precise lesion localization. We conducted an evaluation of our method using the ISIC2016 and ISIC2017 datasets, which yielded findings that demonstrate an AUC of 86.3% for skin lesion classification for ISIC2016 and an average AUC of 92.9% for ISIC2017. Furthermore, our system achieved diagnostic results of significant reference value, with an average AUC of 92% when tested on real-life skin. The experimental results underscore the portable device’s capacity to provide reliable diagnostic information for potential skin lesions, thereby demonstrating its practical applicability.