Subject:
Image segmentation is the process of automatically contouring structures of interest in an image. It has a lot of practical applications for medical studies. We will focus on recent approaches involving the combination of atlas based segmentation and machine learning. The main idea is to learn a local segmentation function in each voxel of an image of support vector machine (SVM) classifiers and use it for the atlas based segmentation of a new image.
The main limitation of this approach is the computational time of the training step: millions of classifiers (one per voxel) need to be trained. The goal of this internship is to use the specificities of the problem to reduce drastically the training time of the SVMs as well as the time required for the optimization of the SVM parameters. An interesting lead based on the learning of solution path has been explored and seem promising. It will be developed during the intership.
Objectives:
- optimizing the implementation of a SVM solution path learning algorithm
- apply this algorithm for the segmentation of 3D MR images
- development and implementation of strategies and algorithms for SVM parameters optimization
Requirements:
- very good scientific background in applied mathematics, numerical analysis
- very good programming skills, experience in C++ is appreciated
- very good knowledge in machine learning more specifically SVM
- knowledge in image segmentation, image registration and general image processing methods is appreciated
Keywords: atlas based segmentation, support vector machine, solution path learning
Reference: Sdika M., Enhancing Atlas Based Segmentation with Multi-Class Linear Classifiers, Medical Physics, In press
Where: Campus de la Doua, Villeurbanne
When: 1er semester 2016
Who: michael.sdika@creatis.insa-lyon.fr