Development of a multi-task deep learning system for classification of nine common knee abnormalities on MRI: a...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Orthopedic Diagnostics

Background:

• Knee MRI interpretation is crucial but time-consuming and error-prone.
• Accurate assessment of multiple knee abnormalities is vital for treatment planning.
• A deep learning system (DLS) was developed to address these challenges.

Purpose of the Study:

• To develop a DLS for classifying common knee abnormalities using MRI.
• To validate the DLS's ability to improve radiologists' diagnostic performance.
• To assess the DLS across multiple centers and radiologist experience levels.

Main Methods:

• A multicenter dataset of 14,962 knee MRIs from 13,419 patients was used.
• The DLS employed an attention-guided coarse-to-fine approach for classifying nine knee abnormalities.
• Three validation steps involved internal and external testing, and radiologist performance assessment.

Main Results:

• The DLS achieved high AUCs for classifying primary and secondary knee abnormalities (0.744-0.898).
• DLS performance was comparable to senior radiologists for specific injuries like meniscal tears and ACL tears.
• Radiologists using the DLS showed improved accuracy and reduced reading times.

Conclusions:

• The developed DLS effectively classifies nine distinct knee abnormalities on MRI.
• Integrating the DLS with radiologists enhances diagnostic accuracy and efficiency.
• Further prospective multi-national studies are recommended to evaluate real-world clinical utility.