Ni Ming, Zhao Yuqing, Wen Xiaoyi, Lang Ning, Wang Qizheng, Chen Wen, Zeng Xiangzhu, Yuan Huishu
Department of Radiology, Peking University Third Hospital, Beijing, China.
Institute of Statistics and Big Data, Renmin University of China, Beijing, China.
Quant Imaging Med Surg. 2023 Jan 1;13(1):80-93. doi: 10.21037/qims-22-470. Epub 2022 Oct 13.
The classification of calcaneofibular ligament (CFL) injuries on magnetic resonance imaging (MRI) is time-consuming and subject to substantial interreader variability. This study explores the feasibility of classifying CFL injuries using deep learning methods by comparing them with the classifications of musculoskeletal (MSK) radiologists and further examines image cropping screening and calibration methods.
The imaging data of 1,074 patients who underwent ankle arthroscopy and MRI examinations in our hospital were retrospectively analyzed. According to the arthroscopic findings, patients were divided into normal (class 0, n=475); degeneration, strain, and partial tear (class 1, n=217); and complete tear (class 2, n=382) groups. All patients were divided into training, validation, and test sets at a ratio of 8:1:1. After preprocessing, the images were cropped using Mask region-based convolutional neural network (R-CNN), followed by the application of an attention algorithm for image screening and calibration and the implementation of LeNet-5 for CFL injury classification. The diagnostic effects of the axial, coronal, and combined models were compared, and the best method was selected for outgroup validation. The diagnostic results of the models in the intragroup and outgroup test sets were compared with those results of 4 MSK radiologists of different seniorities.
The mean average precision (mAP) of the Mask R-CNN using the attention algorithm for the left and right image cropping of axial and coronal sequences was 0.90-0.96. The accuracy of LeNet-5 for classifying classes 0-2 was 0.92, 0.93, and 0.92, respectively, for the axial sequences and 0.89, 0.92, and 0.90, respectively, for the coronal sequences. After sequence combination, the classification accuracy for classes 0-2 was 0.95, 0.97, and 0.96, respectively. The mean accuracies of the 4 MSK radiologists in classifying the intragroup test set as classes 0-2 were 0.94, 0.91, 0.86, and 0.85, all of which were significantly different from the model. The mean accuracies of the MSK radiologists in classifying the outgroup test set as classes 0-2 were 0.92, 0.91, 0.87, and 0.85, with the 2 senior MSK radiologists demonstrating similar diagnostic performance to the model and the junior MSK radiologists demonstrating worse accuracy.
Deep learning can be used to classify CFL injuries at similar levels to those of MSK radiologists. Adding an attention algorithm after cropping is helpful for accurately cropping CFL images.
磁共振成像（MRI）上跟腓韧带（CFL）损伤的分类耗时且阅片者间差异较大。本研究通过将深度学习方法与肌肉骨骼（MSK）放射科医生的分类进行比较，探讨使用深度学习方法对CFL损伤进行分类的可行性，并进一步研究图像裁剪筛选和校准方法。
回顾性分析我院1074例行踝关节镜检查和MRI检查患者的影像资料。根据关节镜检查结果，将患者分为正常组（0级，n = 475）；退变、拉伤和部分撕裂组（1级，n = 217）；以及完全撕裂组（2级，n = 382）。所有患者按8:1:1的比例分为训练集、验证集和测试集。预处理后，使用基于区域的掩膜卷积神经网络（R-CNN）裁剪图像，随后应用注意力算法进行图像筛选和校准，并使用LeNet-5进行CFL损伤分类。比较轴向、冠状位和联合模型的诊断效果，选择最佳方法进行组外验证。将模型在组内和组外测试集的诊断结果与4名不同年资的MSK放射科医生的结果进行比较。
使用注意力算法的Mask R-CNN对轴向和冠状位序列左右图像裁剪的平均平均精度（mAP）为0.90 - 0.96。LeNet-5对轴向序列0 - 2级分类的准确率分别为0.92、0.93和0.92，对冠状位序列分别为0.89、0.92和0.90。序列组合后，0 - 2级分类准确率分别为0.95、0.97和0.96。4名MSK放射科医生对组内测试集分类为0 - 2级的平均准确率分别为0.94、0.91、0.86和0.85，均与模型有显著差异。MSK放射科医生对组外测试集分类为0 - 2级的平均准确率分别为0.92、0.91、0.87和0.85，2名资深MSK放射科医生的诊断性能与模型相似，初级MSK放射科医生的准确率较低。
深度学习可用于对CFL损伤进行与MSK放射科医生相似水平的分类。裁剪后添加注意力算法有助于准确裁剪CFL图像。
