Automatic segmentation of human knee anatomy by a convolutional neural network applying a 3D MRI protocol

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Radiology
• Orthopedic Imaging

Background:

• Magnetic Resonance (MR) imaging is crucial for knee anatomy visualization.
• Accurate segmentation of knee structures is vital for diagnosis and surgical planning.
• Deep learning offers potential for automated and precise anatomical segmentation.

Purpose of the Study:

• To develop and evaluate a deep learning model for segmenting 13 knee anatomical classes using 3D MR imaging.
• To assess the clinical utility of deep learning-based knee anatomy segmentation.
• To compare the performance of different combinations of MR pulse sequences for segmentation accuracy.

Main Methods:

• A DenseVNet deep learning model was trained on 40 healthy knee MR datasets.
• Four 3D MR pulse sequences (T1 TSE, PD TSE, PD FS TSE, Angio GE) were utilized.
• Five input combinations of these sequences were tested, and segmentation performance was evaluated using Dice Similarity Coefficient (DSC), Jaccard index, and Hausdorff distance.

Main Results:

• Combining all four MR sequences yielded the best segmentation performance across all 13 anatomical classes.
• High Dice Similarity Coefficients (DSCs) were achieved for most structures, including Bone Medulla (0.997), muscle (0.998), and cartilage (0.966).
• The model successfully identified an anterior cruciate ligament (ACL) tear in a test subject, demonstrating potential clinical application.

Conclusions:

• Convolutional neural networks are highly effective for segmenting knee joint anatomy from 3D MR sequences.
• The demonstrated deep learning model enables automated segmentation, facilitating the creation of 3D models and pathology detection.
• This automated segmentation holds significant value for preoperative orthopedic evaluations.