French
English
Representation learning
We investigate the potential of machine learning methods for effective and relevant representation of medical data and, in particular, medical images. Challenges concern the construction and the exploitation of relevant data-representation spaces, with statistical properties, to relate the data samples in this representation space, and/or reconstruct cases in the original data space.
Segmentation, localization, and detection
We address the following challenges: i) integration of domain knowledge into deep learning, such as hard shape constraints to guarantee the anatomical consistency of any outputs, joint localization and segmentation to improve accuracy, pathological-class imbalance via data-augmentation techniques through generative networks; ii) detection and segmentation of large heterogeneous datasets of 3D images via weakly supervised or unsupervised deep learning methods through domain adaptation or transfer learning; iii) enhancement of multi-modality statistical atlas-based methods via state-of-the-art feature extraction techniques, improving their accuracy (data-to-atlas transfer or inter-modality domain adaptation) and confidence (uncertainty estimation).
Motion estimation, registration, and deformation
Three challenges are identified: the need of very large training datasets, the lack of the ground-truth transformation for supervised learning and the difficulties of pathological-image registration. We address these challenges through: i) adapted transformation models taking into account the intrinsic nature of the modalities of interest. In particular, we develop dedicated deep-learning motion estimator trained with realistic synthetic data generated from a physical simulator to integrate intrinsic properties of the studied modality; ii) pathological-data generation methods that will integrate functional and physiological prior knowledge of the considered pathologies. Such approaches will be used to build dedicated datasets composed by healthy and pathological cases with the corresponding references to train pathology-robust registration networks.
Reconstruction
We work on four major imaging modalities: i) ultrasound to achieve high frame rate imaging (near 1500 f/s) with good image quality based on deep learning-based approach; ii) Spectral photon-counting CT to investigate joint reconstruction and decomposition using deep learning techniques to improve the precision of spectral decomposition with reduced noise; iii) Diffusion Tensor MR images (DT-MRI) to improve the reconstruction of tiny structures such as brain white matter pathways by working on the optimization of the acquisition parameters; iv) MRI for 3D blood measurements in large volumes to improve spatial resolution and quantification by optimizing the under-sampling strategies in terms of efficiency using machine learning and by investigating dedicated reconstruction techniques based on 6D compressed sensing algorithms that will take advantage of data sparsity simultaneously in the spatial, cardiac, respiratory dimensions, and in the velocity domain.