Semi-supervised learning for automatic segmentation of the knee from MRI with convolutional neural networks

Area of Science:

• Medical Imaging
• Biomechanics
• Orthopedics
• Artificial Intelligence

Background:

• Medical image segmentation is vital for biomechanics and orthopedics but is time-consuming and requires expertise.
• Current segmentation workflows present a bottleneck in research and clinical applications.

Purpose of the Study:

• Develop and evaluate Convolutional Neural Networks (CNNs) for automatic segmentation of Magnetic Resonance Imaging (MRI) data.
• Assess the utility of CNNs for biomechanics research methods, specifically for knee structure segmentation.

Main Methods:

• Trained 2D and 3D CNNs using semi-supervised learning on annotated and unlabeled MRI datasets.
• Implemented Monte Carlo patch sampling for enhanced inference accuracy without additional training steps.
• Evaluated segmentation performance using Intersection-over-Union (IoU), Dice Similarity Coefficient (DSC), and surface error.

Main Results:

• 3D CNNs with Monte Carlo patch sampling achieved high IoU (0.978) and DSC (0.989).
• Median surface error ranged from 0.56 to 0.98 mm, indicating high geometric accuracy.
• Generated meshes from CNN segmentations were FE-ready, demonstrating successful finite element (FE) simulations.

Conclusions:

• CNNs offer a powerful tool to supplement manual segmentation in orthopedics and biomechanics, including FE analysis.
• Semi-supervised learning approaches reduce the need for extensive labeled data, making model development more cost-effective.
• Accelerated FE model development facilitates population-based studies using subject-specific models.