Research on the diagnosis of knee joint MRI meniscal tear based on multi-task learning

YING Meng-jie; WANG Yu-fan; QU Cheng; WU Lei; LIU Xu-dong

doi:10.3969/j.issn.2095-252X.2024.11.004

Abstract：Objective To construct a multi-task learning intelligent diagnostic model MCSNetatt for diagnosing knee MRI meniscus injury,and compare the diagnostic results of the model with those of clinical doctors with or without model assistance to evaluate the actual effectiveness of the model in clinical applications.Methods A retrospective analysis was conducted on the diagnosis and treatment data of 259 patients who underwent knee joint MRI examination and subsequent arthroscopic surgery at the Sixth People's Hospital Affiliated to Shanghai Jiao Tong University School of Medicine from January 2021 to December 2022.The patient data were trained,validated,and tested according to a ratio of 159∶50∶50.The constructed single-task classification model CNetatt,the multi-task learning model MCSNet without added attention module,and the multi-task learning model MCSNetatt with added attention module were used to evaluate the condition of medial and lateral meniscus injuries,respectively.At the same time,two clinical doctors were asked to diagnose medial and lateral meniscus injuries separately with or without model assistance.The study used arthroscopic examination results as the gold standard to compare and analyze the accuracy,precision,recall,F1 score,AUC,and ROC curves of evaluation results.Results The F1 scores for diagnosing medial and lateral meniscus injuries using the MCSNetatt model were 0.930 and 0.900,respectively,which were superior to the CNetatt model and MCSNet model.The accuracy of the MCSNetatt model in diagnosing medial and lateral meniscus injuries was 0.940 and 0.920,respectively,with accuracy of 0.947 and 0.909,both of which were better than those of two clinical doctors.With the assistance of the MCSNetatt model,both doctors had significant improvement in the diagnosis of medial and lateral meniscus injuries.Conclusions The model MCSNetatt based on multi-task learning can assist clinicians to diagnose knee meniscus injuries efficiently and accurately.