CMStatistics 2021: Start Registration
View Submission - CMStatistics
B1577
Title: Model-based clustering of dual-energy CT images for tumor analysis Authors:  Segolene Brivet - McGill University (Canada) [presenting]
Faicel Chamroukhi - Caen University, Lab of Mathematics LMNO (France)
Mark Coates - McGill University (Canada)
Reza Forghani - McGill University (Canada)
Peter Savadjiev - McGill University (Canada)
Abstract: Computed Tomography (CT) scans are commonly used for the evaluation of head and neck cancer but automatic tumor analysis can be challenging on conventional CT. We use an advanced form of CT known as Dual-Energy CT (DECT) or spectral CT. DECT may be viewed as a 4D image of a patient: a 3D body volume over a range of spectral attenuation levels. The latter dimension provides, for each voxel, a decay curve representing energy-dependent changes in attenuation that enables tissue characterization beyond what is possible with conventional CT. We propose a clustering method that uses spectral tissue characteristics to segment the image into areas with consistent contours and high-quality features. The clusters could be used in tumor segmentation or cancer outcome prediction. We construct functional mixture models that specifically integrate spatial context in mixture weights, with mixture component densities being constructed upon the energy decay curves as functional observations. This accommodates the spectral energy curve nature of the data. We design unsupervised clustering algorithms by developing dedicated expectation-maximization (EM) algorithms to estimate the maximum likelihood of the model parameters. The method was evaluated on 90 head and neck DECT scans, each containing a tumor contoured by radiologists. Our algorithm performs well in clustering the anatomical tumor region, as demonstrated by comparing its coverage with the ground truth contour.