MR Protocol Optimization With Deep Learning: A Proof of Concept.

<bold>Purpose: </bold>This study was performed to demonstrate that a properly trained convolutional neural net (CNN) can provide an acceptable surrogate for human readers when performing a protocol optimization study. Tears of the anterior cruciate ligament (ACL) were used as a proof of concept for this study.<bold>Methods: </bold>Following institutional review board approval, a curated set of 2007 paired knee MR images was extracted from the author's picture archival and communications system for 1523 normal knees and 484 knees with torn ACLs. A pair (1 fat-saturated (FS) and 1 non-fat-saturated (NFS)) of midline sagittal images was extracted from each knee. CNNs were trained for both the FS and NFS image sets and used to make predictions on a previously unseen test set of images.<bold>Results: </bold>Receiver operating characteristic area under the curve for the NFS and FS CNNs were, respectively, 0.9983 and 0.9988. Specificity was identical (0.993) for both NFS and FS images. FS sensitivity (0.98) and NFS sensitivity (0.88) were statistically significantly different (P = 0.0253).<bold>Conclusions: </bold>Both FS and NFS performed very well for the diagnosis of ACL tears, although FS sensitivity was superior to NFS sensitivity. The CNNs provided an acceptable surrogate for a human reader in this study. Pulse sequence optimization studies such as this can be opportunistically performed on image sets collected for many other machine learning purposes. [ABSTRACT FROM AUTHOR]