3D Vascular Decomposition and Classification for Computer-Aided Detection.

Area of Science:

• Medical Imaging and Analysis
• Computational Geometry
• Machine Learning for Healthcare

Background:

• Accurate vascular decomposition and classification are crucial for computer-aided detection (CADe) and analysis of cerebrovascular diseases.
• Existing methods may struggle with the complexity and variability of vascular structures, particularly in the presence of aneurysms.

Purpose of the Study:

• To introduce a novel weighted approximate convex decomposition (WACD) and classification methodology for enhanced CADe.
• To evaluate the efficacy of WACD in decomposing vascular sections and classifying them using selected eigenvalues.
• To validate the method's performance against expert segmentations for aneurysm analysis.

Main Methods:

• Vascular decomposition framed as a cluster optimization problem, employing a compact geometric decomposition methodology.
• Classification of decomposed vessel sections using relevant eigenvalues derived from recursive feature elimination for optimal feature selection.
• Validation on a dataset of 98 aneurysms in 112 patients, including a longitudinal study of four internal cerebral aneurysm cases.

Main Results:

• The WACD method achieved an estimated 81.5% accuracy in decomposing vessel sections.
• Recursive feature elimination identified a compact subset of eigenvalues, minimizing classification error and improving precision.
• WACD-classified aneurysm sections demonstrated volumetric and surface area comparisons close to expert segmentations.

Conclusions:

• The proposed WACD methodology offers a robust approach for vascular decomposition and classification in CADe.
• The method shows significant potential for accurately detecting and analyzing changes in aneurysm volumes and surface areas.
• WACD provides a valuable tool for quantitative analysis in cerebrovascular research and clinical applications.