Automatic knee cartilage and bone segmentation using multi-stage convolutional neural networks: data from the...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Biomedical engineering

Background:

• Accurate knee cartilage and bone segmentation is crucial for medical research and clinical practice.
• Existing segmentation methods may lack efficiency or accuracy, particularly across diverse datasets.
• Magnetic Resonance Imaging (MRI) is a key modality for visualizing knee joint structures.

Purpose of the Study:

• To evaluate a novel multi-stage convolutional neural network (CNN) framework for automated knee MRI segmentation.
• To assess the accuracy and efficiency of the framework for segmenting cartilage and bone tissues.
• To validate the framework's performance on datasets from the Osteoarthritis Initiative (OAI) and healthy subjects.

Main Methods:

• A two-stage CNN framework was employed for segmentation.
• Stage 1 produced voxel-wise probabilities for cartilage, bone, and background classes.
• Stage 2 refined segmentation using sub-volume analysis incorporating Stage 1 outputs and raw MRI data, validated via 6-fold cross-validation.

Main Results:

• The framework achieved high Dice similarity coefficients for cartilage segmentation on the OAI dataset (e.g., femoral 0.907, medial tibial 0.876).
• Excellent segmentation accuracies were observed for healthy knee cartilage (e.g., femoral 0.938, patellar 0.955).
• Segmentation was efficient, averaging 91 ± 11 seconds per knee, with average surface distances below in-plane resolution.

Conclusions:

• The developed multi-stage CNN framework effectively automates knee cartilage and bone segmentation from MRI.
• The system demonstrates high accuracy and efficiency across varying disease severities and MRI sequences.
• This automated approach facilitates precise quantification for basic science, clinical trials, and patient care.