Friday, 14 December 2018

Maximilien RENARD

From LISA

Jump to: navigation, search
Ixm.jpg

Contents

Contact

maxrenar@ulb.ac.be
tel.:+32 2 650 26 91
fax:+32 2 650 22 98

ULB - SLN CP165/57
50, Av. F.Roosevelt
1050 Bruxelles
Belgium

Research

Generation of Regional Intensity Models to Improve Atlas-based Automatic Segmentation

Introduction

Active Shape Models are one type of automatic segmentation methods described by T. Cootes[1] and based on a statistical model of the shape of an object of interest.

This method iterates over two steps to segment the object in a unknown image:

  1. Deforming the current shape according to image features such as edges, specific gradients or any type of expert rule;
  2. Recompute the parameters of the intrinsic statistical shape model (SSM) to match the shape it describes to the current deformed shape;

The SSM is generated from a training set of segmented examples of the object and acts as an guide to constrain the segmented shape in a realistic shape-space.

Limitations and possible improvements

To ensure good segmentation results, expert rules are often used and give good results[2]. However, new rules have to be defined for each new object or image modality to be segmented which is a tedious process.

T. Cootes proposed to automatically extract a point-wise intensity model to solve this issue[1]. Neighbourhood relations were not considered though and there was a strong dependence to the accurate correspondence between shapes.

Our solution to these problems is to create a spatial-coherence-aware per-region intensity model where each region of the shape would have an associated characteristic intensity signature.
Clustersandintensityprofiles.png

Proposed Algorithm

In order to extract regions where a characteristic intensity signature can be found, we analysed a clustering process described by F. Chung[3] and based on a modified version of the EM clustering algorithm to take neighbourhood relations into account. It consists of three subsequent steps.

First, an intensity signature (e.g. an intensity profile sampled along the normal to a point of the shape) is extracted, for each point of the shape, from each image in the training set.

Then, the signatures from each image are clustered separately using the above-mentioned algorithm.

Finally, all the clustering solutions are merged to extract reliable clusters (found in a large number of clusterings) thus leading to the desired automatically generated intensity model.

Collaborations

This thesis will be conducted in collaboration with the Medical Planing group of the Konrad-Zuse-Zentrum für Informationstechnik Berlin[4] and the Orthopedy-Traumatology department of the Erasme Hospital[5].

References

  1. 1.0 1.1 T. F. Cootes et al.. Statistical Models of Appearance for Computer Vision, 2004.
  2. H. Seim et al.. Model-based auto-segmentation of knee bones and cartilage in MRI data, Proc. MICCAI Workshop Medical Image Analysis for the Clinic: A grand Challenge, p. 215-223, 2010.
  3. F. Chung et al.. Multimodal prior appearance models based on regional clustering of intensity profiles. MICCAI 2009, vol. 5762 Lecture Notes in Computer Science, p. 1051-1058, Springer Berlin / Heidelberg, 2009
  4. http://www.zib.de/
  5. http://www.erasme.ulb.ac.be

Publications