Dense Multi-Scale Graph Convolutional Network for Knee Joint Cartilage Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate segmentation of knee joint cartilage from MR images is crucial for diagnosing osteoarthritis.
• Existing methods often struggle with complex anatomical structures and variations in image quality.

Purpose of the Study:

• To propose a novel Dense Multi-scale Adaptive Graph Convolutional Network (DMA-GCN) for automatic knee joint cartilage segmentation.
• To evaluate the performance of DMA-GCN against traditional and deep learning-based methods using the Osteoarthritis Initiative (OAI) cohort.

Main Methods:

• The DMA-GCN integrates local and global learning through alternating or sequential combinations of convolutional units.
• A densely connected architecture with residual skip connections enables deeper network structures for enhanced feature representation.
• An adaptive graph learning mechanism allows automatic learning of graph structures during training.

Main Results:

• DMA-GCN achieved superior performance across all evaluation metrics compared to competing methods.
• The method demonstrated high segmentation accuracy, with Dice Similarity Coefficients (DSC) of 95.71% for femoral cartilage and 94.02% for tibial cartilage.
• Thorough experimental analysis investigated the impact of various factors on classification rates.

Conclusions:

• The proposed DMA-GCN method offers a significant advancement in automatic knee cartilage segmentation from MR images.
• DMA-GCN's ability to integrate multi-scale local and global information, coupled with adaptive graph learning, leads to state-of-the-art performance.