A²Seg: Adversarial Learning with Channel-Wise Attention for Skin Lesion Segmentation

In the realm of medical image analysis, the segmentation of skin lesions plays a pivotal role in the early detection and treatment of skin-related diseases. Neural network models have been extensively proposed for automatic segmentation task, but suffer from poor segmentation accuracy caused by gradient disappearance. This paper presents a novel adversarial learning model empowered by a channel-wise attention mechanism, called A<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>Seg, specifically tailored for the segmentation of skin lesions. A<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>Seg operates through a dual-network architecture comprising a segmentation network and a critic network, both trained under an adversarial paradigm. The segmentation network is tasked with generating precise label maps, while the critic network evaluates the segmentation quality, promoting a competitive yet collaborative training environment. The model employs a multi-scale L2 loss function, encouraging the extraction of hierarchical features that capture both local and global image dependencies. Our experimental results, obtained from the ISIC-2017 dataset, validate the efficacy of A<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>Seg, showcasing its superior performance in skin lesion segmentation tasks. The findings underscore the potential of adversarial training and attention mechanisms in enhancing medical image segmentation, setting a new benchmark for future research in the field.