基于深度学习的膝关节软骨MR图像分割方法研究

Research on MR Image Segmentation of Knee Cartilage Based on Deep Learning

摘要: 为实现膝关节骨关节炎(KOA)的早期诊断，提高深度学习模型在膝关节软骨MR图像的分割精度，改善模型针对小目标分割效果不理想的不足，基于深度学习，提出一种端到端的EASU-Net。以深度可分离卷积模块代替卷积模块作为基本模块，减少参数量，增加对深层信息的提取。利用基于ECA的金字塔模块获取不同的感受野，克服了U-Net模型单一感受野的局限性，提高了对不同大小目标的分割能力。设计多尺度输出融合的深监督模块，高质量地提取软骨的细节信息。在OAI-ZIB数据集上测试，相比于基本U-Net和其他现有模型，所提方法在膝关节股骨软骨、胫骨软骨的分割方面都取得了更高的精度。

Abstract: In order to realize the early diagnosis of knee osteoarthritis (KOA)，improve the segmentation accuracy of knee MR Cartilage image by deep learning model, and improve themodel’s unsatisfactory segmentation effect on small targets, an end-to-end EAS U-Net was proposed based on deep learning. The depth separable convolutional module is used to replace the convolutional module as the basic module, reducing the number of parameters and increasing the extraction of deep information. The pyramid module based on ECA is used to obtain different receptive fields, which overcomes the limitation of single receptive field of U-Net model and improves the segmentation ability of objects of different sizes. A deep supervision module with multi-scale output fusion was designed to extract detailed cartilage information with high quality. When tested on the OAI-ZIB dataset, the proposed method achieves higher accuracy in the segmentation of femoral and tibial cartilage of the knee compared with the basic U-Net and other existing models.