Multiscale lightweight 3D segmentation algorithm with attention mechanism: Brain tumor image segmentation

This study proposes a lightweight automatic 3D algorithm with an attention mechanism for the segmentation of brain-tumor images to address the challenges. Accurate segmentation of brain-tumor regions in medical images is essential in patient diagnosis using the 3D U-Net; however, existed 3D networks lack the ability to automatically focus on smaller targets and have low precision on segmenting brain-tumor regions. Existed 3D networks use numerous parameters and are difficult to deploy in practice. The 3D-UNet serves as the basic structure of the proposed approach. (1) this study replaces the standard convolutions with hierarchical decoupled convolutions to reduce the number of parameters; (2) this study adds dilated convolutions to enhance the ability of the network to express multiscale information in the bottom convolution module; (3) this study introduces an attention mechanism to the output layer so that the network can automatically focus on the region of the tumor and use the relationships among the enhancing tumor, whole tumor, and tumor core to increase segmentation precision. The results of the proposed network model, applied to the BraTS 2019 test set, revealed that the Dice coefficients of the proposed ADHDC-Net for enhancing tumor, whole tumor, and tumor core were 77.91 %, 89.94 %, and 83.89 %, respectively, using 0.30 M parameters and 25.81 G floating-point operations. Extensive experiments on the BraTS 2018, BraTS 2019 and BraTS 2020 dadasets show that the proposed model has better potential with regard to the efficient segmentation of small brain-tumor regions.