Abstract:
Segmentation of skin cancer from digital images is crucial for accurate disease diagnosis and
support system for computer-aided diagnosis (CAD) on mobile platforms. However, the irreg ular shapes of the lesions, the absence of sharp edges, the existence of artefacts such as hair
follicles and marker colour make this task difficult. Currently, melanoma segmentation is com monly based on fully connected networks (FCNs) and U-Nets. However, the increasing depth
of these neural network models exposes them to inherent issues such as the vanishing gradient
problem and redundant parameters, potentially leading to reduced segmentation performance.
To address these challenges, we propose a novel lightweight network specifically designed for
skin lesion segmentation, featuring a minimal number of learnable parameters (only 0.8 mil lion). Our approach aims to optimize segmentation accuracy while ensuring efficient computa tion, thereby enhancing the precision and scalability of melanoma detection in dermatological
care. Our proposed lightweight Segmentation with Self-Awareness, Spatial Attention, and Focal
Modulation (LSSF) Network adopts an encoder-decoder architecture enhanced with several key
components, including conformer-based focal modulation attention, self-aware attention mech anisms, global spatial attention, and channel shuffle. These elements are meticulously integrated
to optimize the network’s performance in skin lesion segmentation tasks. To assess the efficacy
of our approach, we conducted comprehensive evaluations using four widely recognized bench mark datasets: ISIC 2016, ISIC 2017, ISIC 2018, and PH2. The empirical findings validate
the superior performance of LSSF-Net, particularly evidenced by its exceptional Jaccard index
scores. These results underscore the effectiveness of our network in achieving state-of-the-art
segmentation accuracy across diverse datasets, thereby showcasing its potential as a valuable
tool for dermatological analysis and melanoma detection.
