Fig. 4.

Cross-validation scheme used to assess MIMoSA performance on T1 lesions (T1L) and T2 lesions (T2L) is pictured. Subjects were randomized to either the training set or the testing set. The MIMoSA model was fit using subjects in the training set. To identify the optimal threshold, probability maps were generated for subjects. These maps were thresholded along a grid selected from 0% to 100% by 1% and then the Sørensen-Dice coefficient (DSC) was calculated. The threshold that resulted in the maximum DSC across subjects in the training set was applied as the threshold in the test set. This procedure was iterated 100 times. Summary statistics are based only on the test set data. The same analysis was repeated using OASIS as the segmentation approach.