Multi-Scale Lightweight 3D MRI Brain Tumor Segmentation Algorithm Based on Attention Mechanism

Abstract

This study proposes a lightweight multi-scale 3D MRI brain tumor segmentation algorithm based on an attention mechanism. The aim is to address two major challenges faced by current three-dimensional convolutional neural networks (3D CNNs) in practical applications: the limited ability to capture shallow features and the low deployment efficiency caused by excessive model parameters. Building upon the 3D U-Net architecture, this study introduces three innovative improvements: (1) In the encoder, conventional convolutions are replaced with a multi-scale feature fusion attention mechanism, enhancing the network's capability to capture shallow information. (2) In the decoder, a residual decoupled convolution is incorporated to reduce the number of parameters while preserving rich multi-scale feature representations, thus improving segmentation performance. (3) In the output layer, a hierarchical attention fusion module is proposed to integrate multi-channel information, which strengthens tumor boundaries and local details, further enhancing segmentation accuracy. The proposed model contains only 3.5 million parameters and requires 27.58 GFLOPs for computation. Extensive experiments conducted on the BraTS 2018 and BraTS 2020 datasets demonstrate significant performance improvements. In BraTS 2018, the DSC scores for the enhanced tumor, the whole tumor, and the tumor core reached 80.34%, 90.12%, and 84.77%, respectively; while in BraTS 2020, the corresponding scores were 79.14%, 88.22%, and 85.79%. These results indicate that the proposed model has strong potential for practical application and demonstrates excellent clinical feasibility as a lightweight solution for brain tumor segmentation.