MSML-AttUNet: A hierarchical attention network with multi-scale and multi-task for precision liver tumor segmentation

Liver cancer is a prevalent malignant tumor with high global incidence and mortality rates. Accurate segmentation of liver tumors is crucial for clinical diagnosis and treatment. In this study, we propose a novel network architecture called MSML-AttUNet, which achieves accurate segmentation and detection of liver tumors, fulfilling the diverse clinical demands. The network utilizes multi-layer dilated convolutions for multi-scale feature extraction, extending the receptive field of the convolutional kernel to encompass a wider scope of contextual information. To address diverse clinical needs, MSML-AttUNet introduces a multi-task learning strategy. On top of tumor segmentation, the network employs this strategy to predict multi-scale tumor boundaries, and detect the presence of tumors in the images. Simultaneously, we design a weighted joint loss function to balance the importance of different tasks and facilitate more precise localization and segmentation of tumor regions. Finally we upgrade the classical attention mechanism to the CBAM (Convolutional Block Attention Module) dual attention mechanism. CBAM can model both spatial and channel dimensions simultaneously, better capturing crucial features in the images. Experimental results demonstrate the superior performance of MSML-AttUNet in liver tumor segmentation tasks, achieving a Dice coefficient of 0.8774. This study presents a novel approach for liver tumor segmentation and detection in clinical settings, offering broad application prospects.