MSRGA-Net：用于3D医学图像分割的多尺度混合网络

MSRGA-Net: A Multi-Scale Hybrid Network for 3D Medical Image Segmentation

摘要: 准确的自动化医学图像分割在自动化诊断和医疗中扮演着重要角色。尽管深度学习方法已经被广泛应用于医学图像分割并获得了成功，但卷积神经网络因其局部卷积特性难以有效捕获全局信息，对不同尺寸大小的组织器官缺乏多尺度特征融合；而基于Transformer的先进方法在捕获全局特征信息方面表现良好，但忽略了像素级空间细节，导致边界定位不够清晰。为了应对这些挑战，本文提出了MSRGA-Net，一个创新的分割框架，能够有效融合低级细节信息、长距离依赖关系和跨尺度信息特征。该网络设计中包含四个核心模块：MSConv，通过多尺度卷积提取细粒度特征；BGAFM，一种新颖的注意力融合模块，可动态赋权特征贡献，以线性复杂度感受全局空间信息；RHAM，用于优化多角度空间与语义信息，同时保留深度方向上的重要信息。此外，提出的MSRGB策略通过选择性地融合多尺度特征，增强了空间一致性，减少特征转换过程中的信息丢失。来自Synapse和医疗分割十项全能(MSD)脑肿瘤分割(BraTS)数据集的广泛实验表明，MSRGA-Net在视觉定量分析和客观评估方面优于最先进的方法。

Abstract: Accurate automated medical image segmentation is essential for computer-aided diagnosis and clinical decision-making, yet achieving both precise boundary localization and robust global understanding remains challenging. Although deep learning methods have achieved strong progress, convolutional neural networks still struggle to capture global context because their receptive fields are inherently local, and they often lack effective multi-scale feature fusion for organs and tissues with large variations in shape and size. Transformer-based approaches can model long-range dependencies more effectively, but they usually overlook pixel-level spatial details, which results in blurred boundaries and incomplete structural representations. To overcome these limitations, we propose MSRGA-Net, an efficient segmentation framework designed to integrate low-level spatial details, long-range contextual relations, and cross-scale features in a balanced manner. The network introduces four core components. MSConv extracts fine-grained representations using multi-scale convolution kernels. BGAFM, a block-grid attention fusion module, dynamically adjusts feature contributions and captures global spatial information with linear computational complexity. RHAM further enhances spatial and semantic cues from multiple perspectives and preserves crucial depth-wise information. In addition, the MSRGB strategy selectively aggregates multi-scale features to improve spatial coherence and reduce information loss during feature transitions. These modules work together to maintain detailed boundaries while ensuring strong global understanding. Extensive experiments on the Synapse dataset and the Medical Segmentation Decathlon (MSD) BraTS dataset demonstrate that MSRGA-Net outperforms state-of-the-art methods in both quantitative visual analysis and objective evaluation.