An lightweight multimodal segmentation network for white matter tracts

The brain's white matter tracts are significant for preoperative planning and postoperative prediction of brain surgery, as their lesions can be observed to diagnose the condition and promptly develop treatment plans. Magnetic Resonance Imaging (MRI) is the mainstream brain imaging method, and T1-weighted (T1w) images generated by scanning are often used to map white matter tracts. However, most of the current brain white matter tract studies are limited to mapping white matter tract trajectories, and little is known about the white matter tract's volume, making it difficult for physicians to make a correct judgment. Based on the above problems, we propose a multimodal lightweight 3D end-to-end framework Attention Vnet-Tiny (AVnet-T) to segment brain white matter tracts accurately. AVnet-T (1) parallelizes two different-sized 3D convolution operations at the input layer, enabling more comprehensive extraction of white matter tract features in multi-modal imaging; (2) introduces compression and activation attention in the encoder part to dynamically adjust the weights to ensure that fiber flow features are not lost in compressing resolution information and aggregating semantic information; (3) significantly reduces the number of its parameters and speeds up its convergence speed when maintaining the performance. We used the BrainPTM2021 challenge online validation tool to evaluate our model with an average dice coefficient of 0.7072. The experimental results demonstrate that our proposed model is efficient and lightweight in automating the segmentation of brain white matter tracts.