Loppini Mattia, Gambaro Francesco Manlio, Chiappetta Katia, Grappiolo Guido, Bianchi Anna Maria, Corino Valentina D A
Department of Biomedical Sciences, Humanitas University, Via Rita Levi Montalcini 4, 20090 Pieve Emanuele, MI, Italy.
IRCCS Humanitas Research Hospital, Via Alessandro Manzoni 56, 20089 Rozzano, MI, Italy.
Bioengineering (Basel). 2022 Jun 29;9(7):288. doi: 10.3390/bioengineering9070288.
Background: Total hip arthroplasty (THA) follow-up is conventionally conducted with serial X-ray imaging in order to ensure the early identification of implant failure. The purpose of this study is to develop an automated radiographic failure detection system. Methods: 630 patients with THA were included in the study, two thirds of which needed total or partial revision for prosthetic loosening. The analysis is based on one antero-posterior and one lateral radiographic view obtained from each patient during routine post-surgery follow-up. After pre-processing for proper standardization, images were analyzed through a convolutional neural network (the DenseNet169 network), aiming to predict prosthesis failure. The entire dataset was divided in three subsets: training, validation, and test. These contained transfer learning and fine-tuning algorithms, based on the training dataset, and were implemented to adapt the DenseNet169 network to the specific data and clinical problem. Results: After the training procedures, in the test set, the classification accuracy was 0.97, the sensitivity 0.97, the specificity 0.97, and the ROC AUC was 0.99. Only five images were incorrectly classified. Seventy-four images were classified as failed, and eighty as non-failed with a probability >0.999. Conclusion: The proposed deep learning procedure can detect the loosening of the hip prosthesis with a very high degree of precision.
全髋关节置换术（THA）的随访传统上通过系列X线成像进行，以确保早期识别植入物失败。本研究的目的是开发一种自动化的放射学失败检测系统。方法：630例THA患者纳入研究，其中三分之二因假体松动需要全部或部分翻修。分析基于每位患者在术后常规随访期间获得的一张前后位和一张侧位X线片。在进行适当标准化的预处理后，通过卷积神经网络（DenseNet169网络）对图像进行分析，旨在预测假体失败。整个数据集分为三个子集：训练集、验证集和测试集。这些子集包含基于训练数据集的迁移学习和微调算法，并被用于使DenseNet169网络适应特定数据和临床问题。结果：经过训练程序后，在测试集中，分类准确率为0.97，敏感性为0.97，特异性为0.97，ROC曲线下面积为0.99。只有五张图像被错误分类。74张图像被分类为失败，80张图像被分类为未失败，概率>0.999。结论：所提出的深度学习程序能够以非常高的精度检测髋关节假体的松动。
