Automated skin lesion segmentation using attention-based deep convolutional neural network

Abstract

Edge detection for dermoscopic images has always been a crucial task for automatic lesion delineation processes. A skin lesion is an area of the skin that takes the form an abnormal growth or appearance when compared to the skin surrounding it. The abnormal appearance is the colored area of the skin that is advised for urgent referral and treatment. The manual way of diagnosing the disease is time-consuming and not quantifiable. However, computer-aided diagnosis (CADx)-based treatment can provide aid to manual delineation by the experts in diagnosing the disease with more proficiency. To advance the digital process of segmentation, a deep learning-based end-to-end framework is proposed for automatic dermoscopic image segmentation. The framework has the modified form of U-Net, which effectively uses Group Normalization (GN) in the encoder and the decoder layers. Attention Gates (AG) focusing on minute details in the skip connection later incorporates with Tversky Loss (TL) as the output loss function are added. Instead of Batch Normalization (BN), GN is used to extract the feature maps generated by the encoding path efficiently. To distinguish high dimensional information from low-level irrelevant background regions in the input image, AGs are used. Tversky Index (TI)-based TL is applied to accomplish better alliance between recall and precision. To further strengthen feature propagation and encourage feature reuse, atrous convolutions are applied in the connecting bridge between the encoder path and the decoder path of the network. The proposed model is evaluated on the ISIC 2018 image dataset, outshone the state-of-the-art segmentation methods.