Title: Hierarchical functional clustering using equivalence test with application to perfusion imaging
Authors: Yves Rozenholc - University Paris Descartes (France) [presenting]
Fuchen Liu - University Paris Descartes (China)
Charles A Cuenod - University Paris Descartes (France)
Abstract: Perfusion imaging allows non invasive access to tissue micro-vascularization. Promising tool to build imaging biomarkers, it suffers from low SNR, improved by averaging homogeneous functional information in large regions of interest. We propose a new automatic segmentation of such image sequence into functionally homogeneous regions. At its core, HiSET (Hierarchical Segmentation using Equivalence Test) aims to cluster functional signals discretely observed with noise on a finite metric space. Assuming independent fixed Gaussian noise, HiSET uses p-values of a multiple equivalence test as dissimilarity measure. It consists of two steps only varying through the neighborhood structure. The first benefit from local regularities on the metric space to control the complexity, the second recovers (spatially) disconnected homogeneous structures at a larger scale. Given a maximal expected homogeneity discrepancy $\delta$, both steps stop automatically through a control of the type I error, providing an adaptive segmentation. Tuning parameter $\delta$ can be interpreted as a multi-resolution diameter around functional patterns recover by the segmentation. When the landscape is functionally piecewise constant with well separated functional features, HiSET is proven to retrieve the exact partition with high probability when the number of observation times is large enough. HiSET outperforms state-of-the-art clustering methods for perfusion imaging sequences.