AM-Net: A Network with Attention and Multi-scale Feature Fusion for Skin Lesion Segmentation

Accurate segmentation of skin lesions is crucial for the diagnosis and treatment of skin diseases. Common problems in dermoscopic medical images, such as inconsistent scale information, blurred small-size boundaries, and irregular shapes of lesion areas, limit the performance of existing methods. To this end, we proposed AM-Net to effectively alleviate the above problems. To address the issue of inconsistent scale information, we designed the multiscale feature integration module (MFIM) and the multiscale feature fusion module (MFFM) as the basic modules for network encoding and decoding. The MFIM integrates the feature information of different scales to enhance feature extraction, and the MFFM processes the multiscale information in parallel to effectively fuse the image features. To address the problem of blurred small-size boundaries, we designed the detail boundary enhancement attention module (DBEAM), which strengthens key details and boundary information in images using an attention-mechanism-weighted approach. To address the problem of irregular shapes of skin lesions, we designed the spatial-channel feature fusion module (SCFFM) to effectively combine feature information at different levels in the encoder-decoder for interaction, enhancing the segmentation capability of irregularly shaped lesion areas. The experiments on the ISIC-2016, ISIC-2017, and $\text {PH}^{{2}}$ datasets demonstrate that our method achieves Dice coefficients of 0.9329, 0.873, and 0.9149, respectively, outperforming existing advanced methods and effectively achieving precise segmentation of dermoscopic image lesions. Our code is available at https://github.com/8yike/AM-Net.git.