VALUE OF DEEP CONVOLUTIONAL NEURAL NETWORK BASED ON SINGLE MRI IMAGES IN DIAGNOSIS OF ANTERIOR CRUCIATE LIGAMENT TEARS OF THE KNEE

Objective To establish a deep convolutional neural network (DCNN) model based on the single MRI image of the knee, and to investigate its value in the diagnosis of anterior cruciate ligament (ACL) tears. Methods Knee MRI images were collected from 1 663 subjects from the GreatPACS image archiving and communication system in No. 971 Hospital of People’s Liberation Army Navy from January 1, 2017 to June 30, 2022, and one image was selected from the MRI images of each patient and was annotated as normal ACL or ACL tears by an orthopedic specialist, which obtained 1 111 images with normal ACL and 552 images with ACL tears. The images were randomly assigned to the training set (1 383 images) and the test set (280 images) at a ratio of 83% and 17%, respectively, to train and test the DCNN model established for the intelligent diagnosis of ACL. The performance of the model was evaluated by positive predictive value (PPV), negative predictive value (NPV), accuracy, sensitivity, specificity, and area under the ROC curve (AUC). Results A DCNN model was successfully established for the intelligent diagnosis of ACL. The test results of this model showed a PPV of 52.99%, an NPV of 88.96%, an accuracy of 73.93%, a sensitivity of 77.50%, and a specificity of 72.50%, with an AUC of 0.602. Conclusion The DCNN model based on single MRI images can help clinicians with the diagnosis of ACL tears.