Beyond pixel: Superpixel-based MRI segmentation through traditional machine learning and graph convolutional network.

Area of Science:

• Medical Imaging Analysis
• Computational Anatomy
• Biomedical Engineering

Background:

• Accurate tendon segmentation is vital for diagnosing pathologies like tendinopathy.
• Automated methods enhance the detailed analysis of specific tendon regions.
• The Achilles tendon, the body's largest tendon, is a key focus for segmentation research.

Purpose of the Study:

• To develop and evaluate an end-to-end module for Achilles tendon segmentation using MRI data.
• To compare the efficacy of superpixel classification (Random Forest, SVM) versus graph-based convolutional networks (GCN) for tendon segmentation.

Main Methods:

• A two-stage approach involving preliminary superpixel-based coarse segmentation followed by final classification.
• Superpixels were classified using Random Forest (RF) and Support Vector Machine (SVM) algorithms.
• An alternative approach converted superpixel arrangements into graphs for classification using a graph-based convolutional network (GCN).

Main Results:

• The RF and SVM approaches achieved high Area Under the ROC Curve (AUC) scores of 0.992 and 0.987, respectively, on unseen test data.
• Sensitivities for RF and SVM were 0.904 and 0.966, demonstrating strong performance in identifying tendon pixels.
• The GCN approach yielded an AUC of 0.933 with a sensitivity of 0.899, indicating good but comparatively lower performance on this dataset.

Conclusions:

• Superpixel generation is an effective coarse segmentation technique for tendon segmentation.
• Non-graph-based superpixel classification methods (RF, SVM) demonstrated superior performance compared to the GCN approach on the limited dataset.
• The study provides valuable insights into tendon segmentation and highlights avenues for future research and model improvement.