Open-box spectral clustering: applications to medical image analysis

Area of Science:

• Computer Vision
• Data Science
• Medical Imaging Analysis

Background:

• Spectral clustering is a powerful technique for 3D image analysis.
• Practical application requires extensive parameter tuning and expert knowledge, especially without labeled data.
• Decisions on cluster number, hierarchical clustering, distance measures, and graph parameters are complex.

Purpose of the Study:

• To simplify the parameter tuning and decision-making process in spectral clustering for 3D image analysis.
• To develop an interactive, open-box system that visualizes mathematical quantities and provides feedback.
• To link the abstract feature space of spectral clustering to the 3D data space for better understanding and prediction.

Main Methods:

• Developed an interactive system for spectral clustering with an open-box approach.
• Visualized mathematical quantities and suggested parameter values with immediate feedback.
• Linked high-dimensional feature space to the 3D data space for improved interpretability.
• Supported various input formats including meshes, grids, and point clouds.
• Enabled outlier filtering and cluster labeling with transferable user actions.

Main Results:

• The system simplifies spectral clustering by providing interactive visualization and parameter suggestions.
• It enhances analyst understanding by connecting feature space to 3D data.
• Developed automated segmentation protocols for chest CT and brain MRI.
• Successfully applied developed protocols to new datasets.

Conclusions:

• The proposed open-box framework significantly simplifies spectral clustering for 3D image analysis.
• The system facilitates better decision-making, improved understanding, and reproducible results.
• It enables the development of robust, automated segmentation protocols for medical imaging.