LMSST-GCN: Longitudinal MRI sub-structural texture guided graph convolution network for improved progression...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Osteoarthritis Research

Background:

• Accurate prediction of knee osteoarthritis (KOA) progression is crucial for personalized interventions.
• Existing methods often focus on single time points or sub-structures, limiting predictive performance.
• Longitudinal analysis of multi-sequence MRI data offers potential for improved KOA progression prediction.

Purpose of the Study:

• To develop and validate a novel longitudinal MRI sub-structural texture-guided graph convolution network (LMSST-GCN) for enhanced KOA progression prediction.
• To compare the performance of LMSST-GCN against traditional clinical and machine learning models.
• To identify key knee sub-structures and their textural changes associated with KOA progression.

Main Methods:

• Utilized longitudinal MRI scans from 600 KOA participants over 24 months.
• Segmented 32 knee sub-structures using 3D nnU-net on IW and DESS sequences.
• Extracted and selected radiomic features, encoded patients into graph representations, and applied EdgeGCN for progression prediction.
• Performed interpretability analysis using GNNExplainer.

Main Results:

• The LMSST-GCN model achieved superior performance (AUC ≥ 0.82) compared to clinical (AUC ≤ 0.72) and machine learning models (AUC ≤ 0.77).
• Highest AUC of 0.85 was obtained using all available longitudinal, multi-sequence MRI data.
• Interpretability analysis highlighted cartilage loss, subchondral bone sclerosis, meniscus injury, and fat pad changes as key indicators of progression.

Conclusions:

• The LMSST-GCN model provides a novel and effective strategy for predicting KOA progression by analyzing longitudinal multi-sequence MRI.
• The model accurately identifies patients at high risk of progression through graph-based vertex classification.
• Publicly available code facilitates further research and application of this method.