23/09/2016 - 14:00
Computer aided diagnosis of epilepsy lesions based on multivariate and multimodality data analysis
Amphithéâtre Claude CHAPPE, INSA Lyon
Type d'Evenements: 

Titre :

Computer aided diagnosis of epilepsy lesions based on multivariate and multimodality data analysis

Date et heure :

Vendredi 23 Septembre 2016 à 14h00

Lieu :

Amphithéâtre Claude CHAPPE, INSA Lyon

Jury :

- RAKOTOMAMONJY Alain, Professeur des Universités, INSA Rouen (rapporteur)

- RUECKERT Daniel, Professeur des Universités, Imperial College London (rapporteur)

- CANU Stéphane, Professeur des Universités, INSA Rouen

- HAMMERS Alexander, Professeur des Universités, King's College London

- FRIBOULET Denis, Professeur des Universités, INSA Lyon

- LARTIZIEN Carole, Chargé de Recherche, CNRS

Ecole Doctorale : EEA



One third of patients suffering from epilepsy are resistant to medication. For these patients, surgical removal of the epileptogenic zone offers the possibility of a cure. Surgery success relies heavily on the accurate localization of the epileptogenic zone. Theanalysis of neuroimaging data such as magnetic resonance imaging (MRI) and positron emission tomography (PET) is increasingly used inthe pre-surgical work-up of patients and may offer an alternative to the invasive reference of Stereo-electro-encephalography (SEEG) monitoring. To assist clinicians in screening these lesions, we developed a computer aided diagnosis system (CAD) based on a multivariate data analysis approach. Ourfirst con- tribution was to formulate the problemof epileptogenic lesion detection as an outlier detection problem. The main motivation forthis formulation was to avoid the dependence on labelled data and the class imbalance inherent to this detection task. The proposed system builds upon the one class support vector machines (OC-SVM) classifier. OC-SVM was trained using features extracted from MRI scans of healthy control subjects, allow- ing a voxelwise assessment of the deviation of a test subject pattern from the learned patterns. System performance was evaluated using realistic simulations of challenging detection tasks as well as clinical data of patients with intractable epilepsy. The out- lier detection framework was further extended to take into account the specificities of neuroimaging data and the detection task at hand. Three original contributions in the domain of outlier detection algorithms were proposed and evaluated based on referenced databases (UCI databases) and on the MRI epilepsy database, when available. First, to handle the presence of noise in the training data, we proposed a reformulation of the sup- port vector data description (SVDD) method, a variant of the OC-SVM method, based on the l0-pseudo-norm. We demonstrated that the resulting l0-SVDD problem can be solved using an iterative procedure providing data specific weighting terms. Second, to deal with the multi-parametric nature of the neuroimaging data, an optimal late fusion strategy for combining multiple base one-class classifiers was investigated. The late fusion approach consisted in building local models associated each with a single MR sequence (DTI, Flair..) and then combining their output based on an original score combination. This approach was shown to outperform standard early fusion approach based on a single global model learned using features extracted from all MR sequences. Finally, tohelp with score interpretation, threshold selection and score combination, we proposed to transform the score outputs of the outlier detection algorithm into well calibrated probabilities. A two steps strategy was proposed. Wefirst generalized the SVDD method by reformulating the associated problem to estimate nestedprobability level-sets and then used a calibra- tion function to convert the outputtedscores into well-calibrated probability estimates. Two calibration functions were considered: the sigmoid function classically used in binary classification problems, and the generalized extreme value distribution, much more suited for long-tailed probability distributions that can be encountered in the context of outlier detection.