Automatic Bone Segmentation from MRI for Real-Time Knee Tracking in Fluoroscopic Imaging.

Area of Science:

• Medical Imaging
• Orthopedics
• Artificial Intelligence in Medicine

Background:

• Real-time knee bone tracking from fluoroscopic imaging aids patellofemoral syndrome research.
• CT imaging for knee kinematics introduces additional ionizing radiation.
• Magnetic Resonance Imaging (MRI) offers a radiation-free alternative for bone segmentation.

Purpose of the Study:

• To develop and validate a deep learning model for segmenting knee bone structures from MRI.
• To improve the geometric accuracy of 3D knee mesh templates derived from MRI.
• To reduce patient exposure to ionizing radiation in knee kinematics studies.

Main Methods:

• A 2.5D U-Net deep neural network architecture was employed for bone segmentation.
• The SKI10 MRI database, augmented with a 'UofA Patella' label, was used for training and testing.
• Segmentation accuracy was initially assessed using Dice scores and Euclidean distance, followed by hyperparameter optimization.

Main Results:

• The 2.5D U-Net achieved a high Dice score of 98% after training.
• Initial segmentation resulted in an unacceptable Euclidean distance (>6 mm) between meshes.
• Hyperparameter optimization, specifically adjusting the output layer threshold, reduced average Euclidean distance to 0.2 mm (variance 0.065 mm).

Conclusions:

• The Dice score alone is insufficient for predicting segmentation geometric accuracy.
• Fine-tuning deep learning model hyperparameters is crucial for achieving high geometric precision in medical imaging.
• This optimized approach provides accurate MRI-derived 3D knee mesh templates for kinematic analysis, minimizing radiation exposure.