A Joint Lightweight U-Shaped Network for Efficient Medical Image Segmentation of Melanoma and Breast Cancer

Abstract

With the continuous development of deep learning, U-Net networks, as an encoder-decoder U-shaped network architecture based on skip connections, have become a popular structure for various medical image segmentation applications in recent years. However, traditional medical segmentation networks face severe challenges when dealing with complex scenarios such as dermoscopy images of melanoma and breast ultrasound images. These challenges primarily stem from limitations in semantic understanding and the complexity of lesion morphology, leading to difficulties in accurately identifying and segmenting lesion structures with irregular shapes and blurred boundaries with surrounding tissues. Additionally, the prevalent issues of parameter redundancy and computational inefficiency in network structures further constrain their potential applications in clinical practice. To address these issues, this paper proposes an image segmentation network based on dynamic skip connections and convolutional multilayer perceptrons—the Joint Lightweight U-shaped Network. JLU-Net, founded on the concept of “joint,” incorporates a joint non-uniform downsampling module that combines linear pooling with nonlinear convolutional downsampling to achieve lightweight modeling. Furthermore, to resolve the semantic gap problem, JLU-Net adopts an enhanced kernel convolution module, which strengthens target region features through feature recalibration operations while integrating detailed and global information. It also includes a joint squeeze attention module, which processes wide and narrow, global and local features simultaneously through squeeze axial operations, thereby enhancing global information exchange. Extensive experiments demonstrate that our JLU-Net achieves state-of-the-art performance across various environments while requiring only 0.29M parameters and 0.52 GFLOPs.