Knee Diagnosis Based On Semi-supervised Learning With Multi-center Magnetic Resonance Images

Knee disorder and abnormality is a common problem that can influence multiple age groups.Magnetic resonance imaging(MRI)can visualize knee’s anatomic structure in a high-resolution manner.So far,some researches have applied deep learning methods to the tasks related with knee MRI and most of them utilize supervised learning(SL)methods.However,SL requires medical experts to annotate every MRI sample which is time-consuming and laborious.By using limited labeled data and abundant unlabeled data,semi-supervised learning(SSL)can effectively mitigate the stress of manual annotation.Additionally,MRI has long scanning time and high cost,which brings about limited MRI data in a single center.But in recent SSL studies,the labeled and the unlabeled are usually ideally separated from the data of a single center,which leads the model to overfit the center’s MRI data and hinders the enhancement of model robustness.Regarding the aforementioned problem,an SSL method which leveraged multicenter MRI data was proposed and dedicated in the knee abnormality classification.Firstly,two knee MRI datasets from different centers were selected as the labeled and the unlabeled.Then the consistency regularization method was utilized to supply the unsupervised loss which only needed the unlabeled data.By combining the unsupervised consistency loss with the supervised classification loss which only utilized the labeled data,the neural network was regularized to extract more discriminative features which benefit knee abnormality classification from MRI images and then these features were used to make a diagnosis on samples.To validate the effectiveness of the proposed SSL method,an SL method was proposed as its counterpart.Finally,some experiments were conducted with these two methods and their corresponding models’ performance were compared under five evaluation metrics.It is indicated that the SSL method is superior to the SL method on both model classification performance and model generalizability on the knee abnormality classification task.Based on the above research,the exponential moving average algorithm was exploited to further upgrade the proposed SSL method.The previous SSL method asked a single model to give the pseudo targets for the unlabeled data,but the improved SSL method integrated the current model and all the previous models into a weighted average model ensemble and the pseudo targets were given by this model ensemble.Firstly,a typical anterior cruciate ligament(ACL)tear dataset and an atypical ACL tear dataset which are from two different centers were chosen as the experimental subjects and then the bidirectional applications of the SSL methods on these two datasets were implemented.The final model performance was evaluated on both the test cohort of the labeled data and the unlabeled data.The experimental results demonstrate that on both data,the previous and the improved SSL methods outperform the SL method and the improved SSL method achieves the best model performance.It suggests that whether the typical or atypical dataset acts as the labeled,the SSL methods can improve the model classification performance on both the labeled data and the unlabeled data.