French
English
Keywords: machine learning, computer science, applied mathematics, atlas based segmentation, image registration, medical image processing, big data
Medical and scientific context: The principle of atlas-based segmentation is to segment an image using the registration to another image with known segmentation. This tool that has many applications in clinical studies. It can be used to measure the size or the atrophy of specific anatomical structures (hippocampus, ...), it also allows to identify the structures for radiotherapy treatment planning or for measuring physiological parameters (fraction anisotropy, T1, T2, ... bone density) in specific structures. The atlas approach has been successfully used to segment a wide variety of organs, on different species and different modalities. This technique is however dependent on the quality of the underlying registration. Results can be dramatically improved if several atlas are used and their respective segmentation are then combined [1]. Drawback of these multi-atlas approaches are mainly the CPU time for the registration to each atlas and the fusion of the segmentation maps. Recently, machine learning techniques have been used to aggregate a whole atlas database information within a single model consisting of an image of local classifiers [2-5]. Only a single registration is required while the results are similar to multi-atlas methods.
Challenges and expected contributions
Challenges: The current deadlock is the time required for the learning on an atlas of local classifiers: one classifier must be trained for each pixel (few millions) in the image using few thousand samples. In the literature, various works based on this approach have therefore presented results with very restricted applications: 2D images, basic local descriptors, limited number of atlas...
Expected contributions: – Development of training algorithms specific for the problem of local classifiers atlas learning. – Investigate among different types of local descriptor of the literature which are the most suited to our problem. – In the laboratory, we already have local linear SVM classifiers kind of atlas [2], but we will test different types of classifiers including those related to deep learning. – The developed methods will be validated on various public databases available covering several potential clinical applications (MRI, US, CT, neuroimaging, cardiac or skeletal muscle).
Profile of the candidate
Very good in: machine learning, applied mathematics, image processing, computer vision, C ++ development.
Very good academic results are required.
Where and When:
Location : CREATIS Lab, La Doua Campus, Villeurbanne, France
Start : October 2016
Contacts :
Michael.Sdika@creatis.insa-lyon.fr
hugues.benoit-cattin@insa-lyon.fr
Bibliography
[1] M. Sdika, “Combining atlas based segmentation and intensity classification with nearest neighbor transform and accuracy weighted vote,” Medical Image Analysis, vol. 14, no. 2, pp. 219 – 226, 2010.
[2] M. Sdika, "Enhancing atlas based segmentation with multiclass linear classifiers.", Medical physics, vol. 42, issue 12, pp. 7169, 2015
[3] Y. Hao, T. Wang, X. Zhang, Y. Duan, C. Yu, T. Jiang, and Y. Fan, “Local label learning (lll) for subcortical structure segmentation: application to hippocampus segmentation,” Human brain mapping, vol. 35, no. 6, pp. 2674–2697, 2014. [4] W. Bai, W. Shi, C. Ledig, and D. Rueckert, “Multi-atlas segmentation with augmented features for cardiac mr images,” Medical image analysis, vol. 19, no. 1, pp. 98–109, 2015.
[5] H. Wang, Y. Cao, and T. Syeda-Mahmood, “Multi-atlas segmentation with learning-based label fusion,” in Machine Learning in Medical Imaging, vol. 8679 of Lecture Notes in Computer Science, pp. 256–263, Springer International Publishing, 2014.