Skin lesion segmentation with a multiscale input fusion U-Net incorporating Res2-SE and pyramid dilated convolution

Skin lesion segmentation is crucial for identifying and diagnosing skin diseases. Accurate segmentation aids in identifying and localizing diseases, monitoring morphological changes, and extracting features for further diagnosis, especially in the early detection of skin cancer. This task is challenging due to the irregularity of skin lesions in dermatoscopic images, significant color variations, boundary blurring, and other complexities. Artifacts like hairs, blood vessels, and air bubbles further complicate automatic segmentation. Inspired by U-Net and its variants, this paper proposes a Multiscale Input Fusion Residual Attention Pyramid Convolution Network (MRP-UNet) for dermoscopic image segmentation. MRP-UNet includes three modules: the Multiscale Input Fusion Module (MIF), Res2-SE Module, and Pyramid Dilated Convolution Module (PDC). The MIF module processes lesions of different sizes and morphologies by fusing input information from various scales. The Res2-SE module integrates Res2Net and SE mechanisms to enhance multi-scale feature extraction. The PDC module captures image information at different receptive fields through pyramid dilated convolution, improving segmentation accuracy. Experiments on ISIC 2016, ISIC 2017, ISIC 2018, PH2, and HAM10000 datasets show that MRP-UNet outperforms other methods. Ablation studies confirm the effectiveness of its main modules. Both quantitative and qualitative analyses demonstrate MRP-UNet's superiority over state-of-the-art methods. MRP-UNet enhances skin lesion segmentation by combining multiscale fusion, residual attention, and pyramid dilated convolution. It achieves higher accuracy across multiple datasets, showing promise for early skin disease diagnosis and improved patient outcomes.