Abstract:
Aiming at the problems of varying sizes, shapes, internal pixel differences, blurry boundaries, and presence of bubbles in skin lesion segmentation, a U-shaped segmentation network, DDAnet, based on dense multi-scale features and dual attention module is proposed. The DenseASPP module in this network obtains rich multi-scale information by densely connecting multiple atrous convolution layers. Meanwhile, the dual attention module composed of CAM and PAM encodes global contextual information to re-register feature maps on channels and positions, achieving emphasis on relevant features and suppression of irrelevant features. These two modules are connected in parallel and work together to improve segmentation performance. On the ISIC2018 dataset, the Acc, JI, DC, Sen and Spec index values of DDAnet are 96.75%, 85.00%, 91.36%, 91.82%, and 97.42%, respectively. The segmentation results are better than those of other segmentation networks, and for extremely challenging cases, DDAnet can still produce accurate and reliable segmentation results, indicating its potential to assist doctors in skin lesion segmentation in clinical diagnosis.