Exploring dual-attention mechanism with multi-scale feature extraction scheme for skin lesion segmentation

Automatic segmentation of skin lesions from dermoscopic images is a challenging task due to the irregular lesion boundaries, poor contrast between the lesion and the background, and the presence of artifacts. In this work, a new convolutional neural network-based approach is proposed for skin lesion segmentation. In this work, a novel multi-scale feature extraction module is proposed for extracting more discriminative features for dealing with the challenges related to complex skin lesions; this module is embedded in the UNet, replacing the convolutional layers in the standard architecture. Further in this work, two different attention mechanisms refine the feature extracted by the encoder and the post-upsampled features. This work was evaluated using the two publicly available datasets, including ISBI2017 and ISIC2018 datasets. The proposed method reported an accuracy, recall, and JSI of 97.5%, 94.29%, 91.16% on the ISBI2017 dataset and 95.92%, 95.37%, 91.52% on the ISIC2018 dataset. It outperformed the existing methods and the top-ranked models in the respective competitions.