TRAINING AND EVALUATION

The reference standard for training the slides selection network was manually selected from the sagittal fat-suppressed T2-weighted fast spin-echo image and sagittal intermediate-weighted fast spin-echo image datasets of all 100 subjects (containing 416 selected image slices from 6126 image slices). The training reference for ligament detection is 416 images labeled by three students who has been trained for more than two months and all labeled slides have been examined by three experienced musculoskeletal radiologist. For ACL tear classification network, a multi-class mask was created for each image slice with the following values: 0=background, 1=with ACL tear, 2=without ACL tear. We also prepare 40 subjects for evaluating the three CNNs using the reference standard interpretation provided by the musculoskeletal radiologist as described previously.