Automatic Deep Learning-assisted Detection and Grading of Abnormalities in Knee MRI Studies

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Orthopedic Diagnostics

Background:

• Knee MRI interpretation for lesions in cartilage, bone marrow, meniscus, and ACL can be subjective.
• Improving interreader agreement is crucial for consistent and reliable diagnosis.

Purpose of the Study:

• To evaluate the efficacy of artificial intelligence (AI) techniques in identifying and assessing knee lesion severity.
• To determine if AI can enhance interreader agreement in knee MRI analysis.

Main Methods:

• Retrospective analysis of 1435 knee MRI studies using a 3D convolutional neural network (CNN).
• AI model developed for detection and grading of abnormalities in cartilage, bone marrow, menisci, and ACL.
• Evaluation of sensitivity, specificity, and Cohen's kappa for AI performance and intergrader agreement on an external dataset.

Main Results:

• AI demonstrated high sensitivity (70-88%) and specificity (85-89%) across all knee tissues.
• Area under the ROC curve ranged from 0.83 to 0.93 for all tissues.
• AI-assisted grading significantly improved intergrader agreement in 10 out of 16 comparisons.

Conclusions:

• 3D CNNs show high accuracy for knee lesion severity scoring.
• AI integration enhances intergrader agreement in knee MRI interpretation, supporting its clinical utility.