DECT-CLUST: Dual-Energy CT Image Clustering and Application to Head and Neck Squamous Cell Carcinoma Segmentation

Area of Science:

• Medical Imaging
• Machine Learning
• Data Science

Background:

• Dual-energy computed tomography (DECT) offers advanced material characterization beyond conventional CT.
• Current head and neck cancer evaluation relies on qualitative assessment and basic measurements.
• DECT provides rich spectral information at each voxel, enabling detailed analysis.

Purpose of the Study:

• To develop novel unsupervised learning techniques for clustering DECT images.
• To integrate spatial context and functional data analysis for DECT image clustering.
• To apply these methods to head and neck squamous cell carcinoma DECT scans.

Main Methods:

• Development of functional mixture models incorporating spatial image context.
• Construction of mixture component densities using DECT energy decay curves as functional observations.
• Implementation of expectation-maximization algorithms for parameter estimation.

Main Results:

• The study presents the first statistical functional data analysis and model-based clustering for DECT spectral information.
• Evaluation on 91 head and neck cancer DECT scans demonstrated the utility of the unsupervised clustering.
• Comparison with manual radiologist contours and baseline algorithms showed promising results.

Conclusions:

• The proposed methodology leverages DECT's full spectral information for advanced image analysis.
• This approach has the potential to enhance downstream machine learning applications for head and neck cancer outcome prediction.
• The methods may add value by providing quantitative insights beyond current clinical evaluations.