DA-SwinUnet：基于方向注意力模块与SwinUnet的股骨头坏死分割模型

DA-SwinUnet: A Segmentation Model for Avascular Necrosis of the Femoral Head Based on Directional Attention Module and SwinUnet

摘要: 股骨头坏死是临床常见的骨科疾病，早期诊断与精准病灶定位对有效治疗至关重要。针对当前影像诊断主观性强、缺乏客观性的问题，以及现有分割方法未能充分利用图像解剖一致性的问题，本研究提出了一种融合定向注意力机制的DA-SwinUnet模型。该模型在Swin-Unet架构基础上引入定向注意力机制以增强方向性特征提取，并通过加权跳跃连接优化网络性能，有效降低了特征融合阶段的计算冗余，提升了对不同大小病灶的分割精度；同时整合方向连通向量概念以进一步强化解剖一致性约束。在股骨头坏死临床MRI数据集上的实验结果显示，DA-SwinUnet模型的Dice系数达到87.3，IoU达到78.7，性能优于传统CNN模型及其他基于Transformer的分割方法。研究表明，DA-SwinUnet能够在保持解剖一致性的前提下实现股骨头坏死病灶的精确分割，显著提升分割性能，为疾病的早期诊断和临床治疗提供了有力支持，具有广阔的前景。

Abstract: Osteonecrosis of the femoral head (ONFH) is a common orthopedic condition, where early diagnosis and precise lesion localization are crucial for effective treatment. To address the issues of subjectivity in current image-based diagnosis and the failure of existing segmentation methods to fully leverage the anatomical consistency of medical images, this study proposes a DA-SwinUnet model enhanced with a directional attention mechanism. Built upon the Swin-Unet Transformer architecture, this model integrates a directional attention mechanism to enhance directional feature extraction and employs weighted skip connections to optimize network performance, reducing computational complexity while improving segmentation accuracy for lesions of varying sizes. The concept of a Directional Connectivity Vector is also incorporated to further enforce anatomical consistency. Experimental results on a clinical MRI dataset for ONFH demonstrate that the DA-SwinUnet model achieves a Dice coefficient of 87.3 and an IoU of 78.7, outperforming traditional CNN models and other Transformer-based segmentation methods. This study indicates that the DA-SwinUnet can achieve precise segmentation of ONFH lesions while maintaining anatomical consistency, significantly enhancing segmentation performance. It provides robust support for the early diagnosis and clinical treatment of the disease and holds promising potential for application.