Knee menisci segmentation using convolutional neural networks: data from the Osteoarthritis Initiative

Area of Science:

• Biomedical Imaging
• Radiology
• Osteoarthritis Research

Background:

• Knee osteoarthritis (OA) diagnosis and progression monitoring are critical.
• Accurate segmentation of menisci from MRI is essential for quantitative biomarker analysis.
• Existing segmentation methods may lack automation and robustness.

Purpose of the Study:

• To introduce a novel automated method for knee menisci segmentation from MRI.
• To evaluate quantitative meniscal biomarkers for OA assessment.
• To determine if these biomarkers can predict incident OA.

Main Methods:

• Developed a segmentation technique using convolutional neural networks and statistical shape models.
• Validated segmentation accuracy against manual segmentations.
• Computed meniscal volume, tibial coverage, and extrusion; tested for OA, joint space narrowing, and pain correlations.
• Assessed biomarker prediction of incident OA using logistic regression.

Main Results:

• Achieved high segmentation accuracy: 83.8% for medial and 88.9% for lateral menisci.
• Identified significantly lower medial tibial coverage and higher medial meniscal extrusion in arthritic knees.
• Found moderate correlation between automated and expert meniscal extrusion measurements.
• Demonstrated medial meniscal extrusion as a significant predictor of incident OA.

Conclusions:

• The automated segmentation method provides excellent accuracy, particularly for medial menisci.
• Validated meniscal biomarkers correlate with OA indicators and expert assessments.
• Medial meniscal extrusion is a confirmed predictor of incident knee osteoarthritis.