Volume 3,Issue 9
基于深度学习的骶髂关节炎CT图像识别方法
CT作为诊断骶髂关节炎病发的重要方式,依赖人工判读,导致误诊和漏诊。为此,本研究提出一种融合2.5D切片输入、多尺度特征提取与融合的深度学习模型MS-2.5D-Net。首先,选取包含骶髂关节中下段关节间隙的连续5层CT切片,构建2.5D输入,在保留三维上下文关联性的同时,降低计算复杂度;其次,修改模型首层卷积适配2.5D输入,在编码阶段嵌入空洞空间金字塔池化(ASPP)模块,同步提取关节间隙的局部微结构特征与全局形态学特征;最后,引入残差特征金字塔(RFPN)跨层融合高层语义与低层细节特征,缓解深层网络梯度消失问题。在29例样本数据集上的实验表明,该模型的敏感度为91.3%,特异性为95.2%,能有效应对骶髂关节炎判别难题,为临床辅助诊断提供了可靠的技术方案。
[1] Salaffi F .Imaging of Sacroiliac Pain: The Current State-of-the-Art[J].Journal of Personalized Medicine, 2024, 14.DOI:10.3390/jpm14080873.
[2]王庆文,曾庆馀,肖征宇,等.磁共振成像对早期骶髂关节炎的诊断价值研究[J].中华风湿病学杂志,2006,10(7):4.DOI:10.3760/j:issn:1007-7480.2006.07.001.
[3] Shenkman Y, Qutteineh B, Joskowicz L, et al. Automatic detection and diagnosis of sacroiliitis in CT scans as incidental findings[J]. Medical image analysis, 2019, 57: 165-175.
[4] Van Den Berghe, Thomas, et al. Neural network algorithm for detection of erosions and ankylosis on CT of the sacroiliac joints: multicentre development and validation of diagnostic accuracy[J]. European radiology 2023,33: 8310-8323.
[5]杜涛,闫建红. DI-MobileNet:基于轻量化网络的骶髂关节炎识别方法[J].智能计算机与应用,2025,15(02):138-143.DOI:10.20169/j.issn.2095-2163.24102205.
[6] Xing Y , Wang J , Zeng G .Malleable 2.5D Convolution: Learning Receptive Fields along the Depth-axis for RGB-D Scene Parsing[J]. 2020.DOI:10.1007/978-3-03058529-7_33.
[7] Tran D , Ray J , Shou Z ,et al.ConvNet Architecture Search for Spatiotemporal Feature Learning[J]. 2017.DOI:10.48550/arXiv.1708.05038.
[8] Xing Y , Wang J , Zeng G .Malleable 2.5D Convolution: Learning Receptive Fields along the Depth-axis for RGB-D Scene Parsing[J]. 2020.DOI:10.1007/978-3-03058529-7_33.
[9] Yang M , Yu K , Zhang C ,et al.DenseASPP for Semantic Segmentation in Street Scenes[J]. 2018.DOI:10.1109/CVPR.2018.00388.
[10]万黎明,张小乾,刘知贵,等.基于空洞空间金字塔池化和多头自注意力的特征提取网络[J].计算机应用,2022,42(S2):79-85.