Department of Orthopedic Surgery, Seoul St. Mary's Hospital, The Catholic University of Korea, Seoul, Republic of Korea.
Am J Sports Med. 2023 Sep;51(11):2824-2830. doi: 10.1177/03635465231189201. Epub 2023 Aug 11.
It is challenging to predict retear after arthroscopic rotator cuff repair (ARCR). The usefulness of arthroscopic intraoperative images as predictors of the ARCR prognosis has not been analyzed.
To evaluate the usefulness of arthroscopic images for the prediction of retear after ARCR using deep learning (DL) algorithms.
Cohort study (Diagnosis); Level of evidence, 2.
In total, 1394 arthroscopic intraoperative images were retrospectively obtained from 580 patients. Repaired tendon integrity was evaluated using magnetic resonance imaging performed within 2 years after surgery. Images obtained immediately after ARCR were included. We used 3 DL architectures to predict retear based on arthroscopic images. Three pretrained DL algorithms (VGG16, DenseNet, and Xception) were used for transfer learning. Training and test sets were split into 8:2. Threefold stratified validation was used to fine-tune the hyperparameters using the training data set. The validation results of each fold were evaluated. The performance of each model in the test set was evaluated in terms of accuracy, area under the receiver operating characteristic curve (AUC), F1-score, sensitivity, and specificity.
In total, 1138 and 256 arthroscopic images were obtained from 514 patients and 66 patients in the nonretear and retear groups, respectively. The mean validation accuracy of each model was 83% for VGG16, 89% for Xception, and 91% for DenseNet. The accuracy for the test set was 76% for VGG16, 87% for Xception, and 91% for DenseNet. The AUC was highest for DenseNet (0.92); it was 0.83 for VGG16 and 0.91 for Xception. For the test set, the specificity and sensitivity were 0.93 and 0.84 for DenseNet, 0.89 and 0.84 for Xception, and 0.70 and 0.80 for VGG16, respectively.
The application of DL algorithms to intraoperative arthroscopic images has demonstrated a high level of accuracy in predicting retear occurrences.
预测关节镜下肩袖修复（ARCR）后的再撕裂是具有挑战性的。关节镜术中图像作为 ARCR 预后预测指标的有用性尚未得到分析。
使用深度学习（DL）算法评估关节镜图像在预测 ARCR 后再撕裂中的作用。
队列研究（诊断）；证据水平，2 级。
总共从 580 名患者中回顾性获得了 1394 个关节镜术中图像。使用手术后 2 年内进行的磁共振成像评估修复肌腱的完整性。包括 ARCR 后立即获得的图像。我们使用 3 种 DL 架构基于关节镜图像来预测再撕裂。使用 3 种预先训练的 DL 算法（VGG16、DenseNet 和 Xception）进行迁移学习。训练集和测试集分为 8:2。使用三折分层验证使用训练数据集调整超参数。评估每个折的验证结果。使用准确性、受试者工作特征曲线（ROC）下面积（AUC）、F1 评分、敏感性和特异性评估每个模型在测试集中的性能。
总共从 514 名患者和 66 名患者中获得了 1138 个和 256 个关节镜图像，分别为无再撕裂组和再撕裂组。每个模型的平均验证准确性为 VGG16 为 83%，Xception 为 89%，DenseNet 为 91%。VGG16 的测试集准确性为 76%，Xception 为 87%，DenseNet 为 91%。DenseNet 的 AUC 最高（0.92）；VGG16 为 0.83，Xception 为 0.91。对于测试集，DenseNet 的特异性和敏感性分别为 0.93 和 0.84，Xception 为 0.89 和 0.84，VGG16 为 0.70 和 0.80。
DL 算法在术中关节镜图像中的应用在预测再撕裂发生方面具有较高的准确性。
