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

Area of Science:

• Medical imaging analysis
• Machine learning in radiology
• Orthopedic diagnostics

Background:

• Accurate diagnosis of anterior cruciate ligament (ACL) tears is crucial for effective patient management.
• Protocol optimization studies in medical imaging are essential for improving diagnostic accuracy and efficiency.
• Convolutional neural networks (CNNs) show promise in medical image analysis tasks.

Purpose of the Study:

• To evaluate a trained convolutional neural network (CNN) as a surrogate for human readers in anterior cruciate ligament (ACL) tear detection.
• To demonstrate the utility of CNNs in protocol optimization studies using knee MRI data.
• To compare the performance of CNNs trained on fat-saturated (FS) and non-fat-saturated (NFS) MRI sequences for ACL tear diagnosis.

Main Methods:

• A dataset of 2007 knee MRI scans (1523 normal, 484 with ACL tears) was curated.
• Midline sagittal images, both FS and NFS, were extracted for each knee.
• CNNs were trained on these image sets and tested on unseen data to predict ACL tears.

Main Results:

• CNNs achieved high diagnostic performance, with areas under the receiver operating characteristic curve of 0.9983 (NFS) and 0.9988 (FS).
• Specificity was excellent and identical (0.993) for both FS and NFS images.
• Fat-saturated (FS) images demonstrated statistically significantly higher sensitivity (0.98) compared to non-fat-saturated (NFS) images (0.88).

Conclusions:

• Both FS and NFS MRI sequences, when analyzed by CNNs, are effective for diagnosing ACL tears.
• Fat-saturated (FS) sequences offer superior sensitivity for ACL tear detection compared to NFS sequences.
• CNNs can serve as acceptable surrogates for human readers in MRI protocol optimization studies, enabling opportunistic research on existing datasets.