浙江医学2025，Vol.47Issue(22)：2409-2415,2424,后插5-后插6,10.DOI:10.12056/j.issn.1006-2785.2025.47.22.2025-1919
基于迁移学习和贝叶斯优化的早期膝骨关节炎深度学习诊断系统的构建
Deep learning diagnostic system establishment for early knee osteoarthritis based on transfer learning and bayesian optimization
摘要
Abstract
Objective To develop a deep learning diagnostic system for early knee osteoarthritis(KOA,KL grades 0-2)based on transfer learning and Bayesian optimization,and to validate the model's performance through external and internal datasets.Methods A total of 8 205 X-ray images of 4 796 patients with anterior and posterior knee joints and standing weight-bearing with both knees were collected from the Osteoarthritis Initiative database between 2004 and 2015.They were partitioned into training,internal validation,and internal test sets at a 70∶15∶15 ratio(5 742,1 229,and 1 234 images respectively).A retrospective collection was made of 246 standing X-ray images of both knee joints from 123 patients with knee pain admitted to the Ningbo Second Hospital from September to December 2024 as the external validation set.The interaction effects between pretraining strategies(RadImageNet,ImageNet,random initialization)and hyperparameter approaches(Bayesian optimization and default configuration)were compared.Model stability was assessed through 10-fold cross-validation,while model reliablity was examined by combining Grad-CAM,t-SNE,calibration curves,and decision curve analysis.External validation dataset was used to evaluate real-world clinical performance of the model.Results The accuracy rate of RadImageNet_Bayesian(RadImageNet pretraining transfer learning and Bayesian optimization combined model)achieved 85.05％accuracy in the internal test set(AUC=0.950 7).The 10-fold cross-validation showed stable performance(AUC=0.950 7±0.023 3).In the external validation cohort,RadImageNet_Bayesian achieved 72.36％accuracy(AUC=0.941 5),representing a 12.69％decrease from the internal test set,falling within the expected range of validation in different clinical sets.Conclusion The deep learning diagnostic system RadImageNet_Bayesian developed in this study demonstrates favorable clinical applicability and cross-domain stability for early KOA detection.However,the difficult identification for KL grade 1 reflects the clinical nature of imaging features at this stage,suggesting the necessity of integrating clinical symptoms for comprehensive diagnosis.关键词
膝关节骨关节炎/深度学习/迁移学习/贝叶斯优化/早期检测Key words
Knee osteoarthritis/Deep learning/Transfer learning/Bayesian optimization/Early detection引用本文复制引用
石林,梁清,杨放,施泽文,庞清江,陈先军..基于迁移学习和贝叶斯优化的早期膝骨关节炎深度学习诊断系统的构建[J].浙江医学,2025,47(22):2409-2415,2424,后插5-后插6,10.基金项目
浙江省省级临床重点专科建设项目(2024021) （2024021）
宁波市骨科与运动康复临床医学研究中心建设项目(2024L004) （2024L004）
