A Multi-Scale Lightweight Brain Glioma Image Segmentation Network

Manual segmentation of brain tumor areas in magnetic resonance imaging (MRI) images is time-consuming and laborious, and it can be easily influenced by individual subjectivity. To reliably and efficiently segment brain tumors semi-automatically or automatically is particularly important for medically assisted diagnosis. In recent years, convolutional neural network-based methods for automatic segmentation of brain tumor images have made great progress, but the existing methods still cannot effectively fuse features in terms of large-scale contours and small-scale texture details of tumor images, and the rich global background information is ignored during training. In view of these problems, this paper proposed a multi-scale lightweight brain tumor image segmentation network MSL-Net. First, the base convolution in the U-Net network was replaced with an improved hierarchical decoupled convolution to expand the perceptual field while efficiently exploring multi-scale multi-view spatial information. Then, a bidirectional feature pyramid network structure was introduced at the skipping connection to fuse multi-scale features, and a hybrid loss function combining the generalized Dice loss function and the Focal loss function was used to improve segmentation accuracy and accelerate convergence in the case of pixel count imbalance between tumor and non-tumor regions. Experimental results on the BraTS 2019 dataset show that the Dice similarity coefficients of the proposed MSL-Net network in the overall tumor region, core tumor region and enhanced tumor region are 0.9003, 0.8306 and 0.7770, respectively, and the number of parameters and computation (floating-point operations per second) are 3.9×10⁵ and 3.16×10¹⁰, respectively. Compared with the current state-of-the-art methods, the method proposed in the paper achieves high segmentation accuracy while achieving light weight.