A Lightweight 3D Brain Tumor Segmentation Network, Dp-Net, Based on Multi-Scale Feature Extraction and Prior Guidance

Brain tumors are one of the most aggressive diseases in the central nervous system, and accurate segmentation is crucial for clinical diagnosis and treatment. Although existing deep learning methods have improved the segmentation accuracy, due to the complexity of tumor morphology, these methods usually rely on complex network structures to improve the accuracy, resulting in excessive computational overhead and difficulty in balancing accuracy and efficiency. To this end, this paper proposes a lightweight 3D segmentation network DP-Net based on multi-scale feature extraction and a priori guidance.DPNet adopts a dynamic multi-scale feature aggregation module based on a hierarchical decoupling architecture instead of the standard convolution, which significantly reduces the number of parameters and at the same time enhances the multi-scale feature extraction capability. In addition, the fusion effect of multi-scale features is further optimized by introducing a twobranch adaptive fusion module. The a priori guided attentionbased fusion module is designed in the output layer, which further enhances the model's ability to recognize small lesions. The experimental results show that the Dice scores of DP-Net in the enhanced tumor (ET), whole tumor (WT), and tumor core (TC) regions are <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$87.3 \%, 91.2 \%$</tex>, and 88.3 %, respectively, and the number of model parameters is only 0.26 M with a computational amount of 8.58 G FLOPS, which proves that it is able to effectively maintain the computational efficiency while improving the segmentation accuracy.