基于ResNet深度网络模型的髋关节置换术后并发症分类研究
Classification of Postoperative Complications after Hip Arthroplasty Based on the ResNet Deep Network Model

目的：分析基于ResNet迁移学习模型鉴别髋关节置换术后假体周围感染与松动的价值。方法：本研究的数据来源于2015年1月至2022年12月期间，在上海第六人民医院骨科接受全髋关节翻修手术的206例患者。这些患者因髋关节置换术后出现假体周围感染或假体松动的情况而接受了翻修手术，收集患者髋关节置换术后X线图像。使用迁移学习方法对髋关节假体周围感染和松动进行鉴别，分别建立ResNet18、ResNet50迁移学习模型，并使用SHAP方法对模型进行可视化分析。结果：通过对两种不同网络的模型进行迁移学习实验对比，得到了以下结果：基于ResNet18网络的迁移学习模型在鉴别髋关节置换术后假体周围感染与松动方面表现出显著优势，模型的准确率达到了91.30%，灵敏度为95.94%，特异度为87.50%，AUC为93.94%。这些指标表明了该模型在对这两种并发症进行区分时的准确性和可靠性。实验还进行了Delong检验，ResNet18网络模型与ResNet50模型之间的AUC差异具有统计学意义(p < 0.05)。结论：本文旨在建立迁移学习诊断模型，为早期临床髋关节置换术后假体周围感染与松动的诊断提供一种方法。
Objective: To analyse the value of identifying periprosthetic infection and loosening after hip arthroplasty based on the ResNet migration learning model. Methods: The data for this study were obtained from 206 patients who underwent total hip revision surgery between January 2015 and December 2022 at the Department of Orthopedics, Shanghai Sixth People’s Hospital. These patients underwent revision surgery due to periprosthetic infection or loosening of the prosthesis after hip arthroplasty, and postoperative X-ray images of the patients’ hip arthroplasty were collected. The migration learning method was used to identify the infection and loosening around the hip prosthesis, and ResNet18 and ResNet50 migration learning models were established respectively, and the models were visualised and analysed using the SHAP method. Results: By comparing the migration learning experiments with the models of two different networks, the following results were obtained: the migration learning model based on the ResNet18 network showed a significant advantage in identifying the infection and loosening around the hip prosthesis after hip arthroplasty, and the accuracy of the model reached 91.30%, the sensitivity was 95.94%, the specificity was 87.50%, and the AUC was 93.94%. These metrics demonstrated the accuracy and reliability of the model in differentiating between these two complications. A Delong test was also performed, and the difference in AUC between the ResNet18 network model and the ResNet50 model was statistically significant (p < 0.05). Conclusion: The aim of this paper is to establish a migratory learning diagnostic model to provide a method for the diagnosis of periprosthetic infection and loosening after early clinical hip arthroplasty.