Center for Bionics, Korea Institute of Science and Technology, Seoul, 02792, Korea.
Department of Orthopedic Surgery, Yeson Hospital, Bucheon, 14555, Korea.
Sci Rep. 2020 Sep 24;10(1):15632. doi: 10.1038/s41598-020-72357-0.
Rotator cuff tear (RCT) is one of the most common shoulder injuries. When diagnosing RCT, skilled orthopedists visually interpret magnetic resonance imaging (MRI) scan data. For automated and accurate diagnosis of RCT, we propose a full 3D convolutional neural network (CNN) based method using deep learning. This 3D CNN automatically diagnoses the presence or absence of an RCT, classifies the tear size, and provides 3D visualization of the tear location. To train the 3D CNN, the Voxception-ResNet (VRN) structure was used. This architecture uses 3D convolution filters, so it is advantageous in extracting information from 3D data compared with 2D-based CNNs or traditional diagnosis methods. MRI data from 2,124 patients were used to train and test the VRN-based 3D CNN. The network is trained to classify RCT into five classes (None, Partial, Small, Medium, Large-to-Massive). A 3D class activation map (CAM) was visualized by volume rendering to show the localization and size information of RCT in 3D. A comparative experiment was performed for the proposed method and clinical experts by using randomly selected 200 test set data, which had been separated from training set. The VRN-based 3D CNN outperformed orthopedists specialized in shoulder and general orthopedists in binary accuracy (92.5% vs. 76.4% and 68.2%), top-1 accuracy (69.0% vs. 45.8% and 30.5%), top-1±1 accuracy (87.5% vs. 79.8% and 71.0%), sensitivity (0.94 vs. 0.86 and 0.90), and specificity (0.90 vs. 0.58 and 0.29). The generated 3D CAM provided effective information regarding the 3D location and size of the tear. Given these results, the proposed method demonstrates the feasibility of artificial intelligence that can assist in clinical RCT diagnosis.
肩袖撕裂（RCT）是最常见的肩部损伤之一。在诊断 RCT 时，熟练的骨科医生会直观地解读磁共振成像（MRI）扫描数据。为了实现 RCT 的自动和准确诊断，我们提出了一种基于深度学习的全 3D 卷积神经网络（CNN）方法。该 3D CNN 可自动诊断 RCT 是否存在、撕裂大小，并提供撕裂位置的 3D 可视化。为了训练 3D CNN，我们使用了 Voxception-ResNet（VRN）结构。该架构使用 3D 卷积滤波器，因此与基于 2D 的 CNN 或传统诊断方法相比，从 3D 数据中提取信息具有优势。使用 2124 名患者的 MRI 数据对基于 VRN 的 3D CNN 进行了训练和测试。该网络的训练目标是将 RCT 分为五种类别（无、部分、小、中、大-巨大）。通过体绘制可视化 3D 类激活图（CAM），以显示 3D 中 RCT 的定位和大小信息。通过使用从训练集中分离出的 200 个随机测试集数据，对所提出的方法和临床专家进行了对比实验。基于 VRN 的 3D CNN 在二进制准确率（92.5%对 76.4%和 68.2%）、准确率（69.0%对 45.8%和 30.5%）、准确率±1（87.5%对 79.8%和 71.0%）、敏感度（0.94 对 0.86 和 0.90）和特异性（0.90 对 0.58 和 0.29）方面均优于专门从事肩部和普通骨科的骨科医生。生成的 3D CAM 提供了有关撕裂 3D 位置和大小的有效信息。鉴于这些结果，该方法证明了人工智能在临床 RCT 诊断中具有辅助作用的可行性。
