Key-words: machine learning, deep learning, generative adversarial networks (GAN), medical imaging
Scientific context:
In recent years, machine learning has received a lot of attention to explore and structure multidimensional and multi-modality medical imaging data for the purpose of image segmentation, registration or automated detection and characterization of pathology. Among the machine learning tools, deep learning is one approach that is currently focusing increasing attention, since it has been recently shown to outperform traditional approaches in most of the applications cited above.
The vast majority of tasks that have been studied so far, in particular with convolutional networks are supervised, meaning that the last layer of the network is a classification or regression layer. Training is thus performed on samples consisting of couples of input (image) and output (label for classification or continuous value for regression) values. Even if patient scans can easily be labelled as pathological at the image level, access to this ground truth at the voxel level is very limited since it is based on the consensus manual annotations of clinical experts which are very time and effort demanding.
One alternative to the lack of labelled data is to learn unsupervised representation models.This topic has been poorly explored in the domain of medical imaging. We recently developed such an approach for neuroimaging applications. We indeed proposed to learn efficient unsupervised representations from patches extracted from MR brain acquisitions of normal subjects based on a siamese network architecture [Alaverdyan et al, 2016]. This latent representation then serves as input to an outlier detection algorithm to detect epileptogenic lesions in epilepsy patient MR scans. Promising results were achieved with such an architecture [Alaverdyan et al, 2017].
Objective
The objective of this project is to further investigate the potential of unsupervised deep medical imaging feature learning. We propose to explore the performance of generative adversarial networks (GAN) architecture to learn more discriminant feature representation than that achieved with the siamese architecture developed in [Alaverdyan, 2016]. We will start by implementing and optimizing a standard GAN architecture and then focus on more complex models that can embed contextual information with the aim to improve the discriminative power of the learned representation. The medical application domain will be the detection of epileptogenic lesions in multiparametric brain magnetic resonance imaging (MRI) [El Azami et al., 2016].
See the attached pdf file for further details
Applications
Interested applicants are required to send a cover letter, CV and any other relevant documents (reference letter, recent transcripts of marks,...) to carole.lartizien@creatis.insa-lyon.fr
References
[El Azami, 2016] El Azami, M., A. Hammers, J. Jung, N. Costes, R. Bouet and C. Lartizien. Detection of Lesions Underlying Intractable Epilepsy on T1-Weighted MRI as an Outlier Detection Problem. PLoS ONE, 11(9): e0161498, 2016.
[Alaverdyan, 2017] Z. Alaverdyan et C. Lartizien. Feature extraction with regularized siamese networks for outlier detection: application to epilepsy lesion detection. Conférence sur l'apprentissage automatique (CAp 2017), Grenoble, France, 2017.
[Alaverdyan, 2016] Z. Alaverdyan, C. Lartizien. Automatic extraction of representations for outlier detection in medical imaging. Réunion du GDR ISIS- thème A : Apprentissage de représentations : méthodologies et applications, Paris, Oct 2016.
Gratuity
~550 euros/mois.