Unsupervised clustering based coronary artery segmentation.

Area of Science:

• Medical Imaging
• Cardiovascular Research
• Computational Anatomy

Background:

• Accurate 3D segmentation of coronary arteries from computed tomography coronary angiography (CTCA) is vital for diagnosing lesions and guiding clinical decisions.
• Manual segmentation is labor-intensive and error-prone, while neural network methods demand extensive data and computational power.
• This work presents an alternative automatic segmentation approach leveraging clustering algorithms and graph structures.

Purpose of the Study:

• To develop and evaluate an automatic coronary artery segmentation methodology.
• To compare the performance of a 2.5D (3Axis) approach with a perpendicular (Perp) cross-sectional method.
• To assess the clinical applicability and explainability of the proposed segmentation technique.

Main Methods:

• An automatic segmentation method combining clustering algorithms and graph structures was developed.
• Two distinct approaches were evaluated: 3Axis (utilizing axial, sagittal, and coronal slices) and Perp (using vessel cross-sections).
• The methods were validated on independent patient datasets, including those with diagnosed lesions.

Main Results:

• The 3Axis method achieved superior performance with Dice scores of 0.88 (test set) and 0.83 (lesion set).
• The Perp method yielded comparable results (0.81 test set, 0.82 lesion set), with slight decreases in lesion regions (0.79-0.80).
• Performance is competitive with existing state-of-the-art automatic segmentation techniques.

Conclusions:

• The clustering-based segmentation framework offers a robust and efficient solution for coronary artery analysis.
• The method enhances explainability through cluster and graph visualization, aiding clinical interpretation.
• This approach shows significant potential for improving accuracy and efficiency in cardiovascular imaging workflows.