# Hierarchical clustering using equivalence test : application on automatic segmentation of dynamic contrast enhanced image sequence

Abstract : Dynamical contrast enhanced (DCE) imaging allows non invasive access to tissue micro-vascularization. It appears as a promising tool to build imaging biomarker for diagnostic, prognosis or anti-angiogenesis treatment monitoring of cancer. However, quantitative analysis of DCE image sequences suffers from low signal to noise ratio (SNR). SNR may be improved by averaging functional information in large regions of interest, which however need to be functionally homogeneous. To achieve SNR improvement, we propose a novel method for automatic segmentation of DCE image sequence into functionally homogeneous regions, called DCE-HiSET. As the core of the proposed method, HiSET (Hierarchical Segmentation using Equivalence Test) aims to cluster functional (e.g. with respect to time) features or signals discretely observed with noise on a finite metric space considered to be a landscape. HiSET assumes independent Gaussian noise with known constant level on the observations. It uses the p-value of a multiple equivalence test as dissimilarity measure and consists of two steps. The first exploits the spatial neighborhood structure to preserve the local property of the metric space, and the second recovers (spatially) disconnected homogeneous structures at a larger (global) scale. Given an expected homogeneity discrepancy $\delta$ for the multiple equivalence test, both steps stop automatically through a control of the type I error, providing an adaptive choice of the number of clusters. Parameter $\delta$ appears as the tuning parameter controlling the size and the complexity of the segmentation. Assuming that the landscape is functionally piecewise constant with well separated functional features, we prove that HiSET will retrieve the exact partition with high probability when the number of observation times is large enough. In the application for DCE image sequence, the assumption is achieved by the modeling of the observed intensity in the sequence through a proper variance stabilization, which depends only on one additional parameter $a$. Therefore, DCE-HiSET is the combination of this DCE imaging modeling step with our statistical core, HiSET. Through a comparison on synthetic 2D DCE image sequence, DCE-HiSET has been proven to outperform other state-of-the-art clustering-based methods. As a clinical application of DCE-HiSET, we proposed a strategy to refine a roughly manually delineated ROI on DCE image sequence, in order to improve the precision at the border of ROIs and the robustness of DCE analysis based on ROIs, while decreasing the delineation time. The automatic refinement strategy is based on the segmentation through DCE-HiSET and a series of erosion-dilation operations. The robustness and efficiency of the proposed strategy are verified by the comparison of the classification of 99 ovarian tumors based on their associated DCE-MR image sequences with the results of biopsy anapathology used as benchmark. Furthermore, DCE-HiSET is also adapted to the segmentation of 3D DCE image sequence through two different strategies with distinct considerations regarding the neighborhood structure cross slices. This PhD thesis has been supported by contract CIFRE of the ANRT (Association Nationale de la Recherche et de la Technologie) with a french company INTRASENSE, which designs, develops and markets medical imaging visualization and analysis solutions including Myrian®. DCE-HiSET has been integrated into Myrian® and tested to be fully functional.
Keywords :
Document type :
Theses
Domain :

Cited literature [97 references]

https://tel.archives-ouvertes.fr/tel-02110810
Contributor : Abes Star <>
Submitted on : Thursday, April 25, 2019 - 3:33:14 PM
Last modification on : Friday, December 20, 2019 - 5:29:24 AM

### File

vd_Liu_Fuchen.pdf
Version validated by the jury (STAR)

### Identifiers

• HAL Id : tel-02110810, version 1

### Citation

Fuchen Liu. Hierarchical clustering using equivalence test : application on automatic segmentation of dynamic contrast enhanced image sequence. Imaging. Université Sorbonne Paris Cité, 2017. English. ⟨NNT : 2017USPCB013⟩. ⟨tel-02110810⟩

Record views