The construction of an atlas of the human brain connectome, in particular, the cartography of fiber bundles of superficial white matter (SWM) is a complex and unachieved task. Its description is essential for the understanding of human brain function and the study of several pathologies. In this work we applied an automatic white matter bundle segmentation method proposed in the literature for the analysis of the variability of a big amount of superficial white matter bundles. The method was applied to 30 subjects of a high quality HARDI database, adding several processing steps in order to improve the results. Then we calculated some indices for studying the variability of 40 SWM fiber bundles from each hemisphere, and we constructed a model of these bundles in the MNI standard space.

Large and performant neural networks are often overparameterized and can be drastically reduced in size and complexity thanks to pruning. Pruning is a group of methods, which seeks to remove redundant or unnecessary weights or groups of weights in a network. These techniques allow the creation of lightweight networks, which are particularly critical in embedded or mobile applications. In this paper, we devise an alternative pruning method that allows extracting effective subnetworks from larger untrained ones. Our method is stochastic and extracts subnetworks by exploring different topologies which are sampled using Gumbel Softmax. The latter is also used to train probability distributions which measure the relevance of weights in the sampled topologies. The resulting subnetworks are further enhanced using a highly efficient rescaling mechanism that reduces training time and improves performances. Extensive experiments conducted on CIFAR10 show the outperformance of our subnetwork extraction method against the related work.

This paper deals with the 2D motion estimation of coronary arteries. It exploits the geometry of acquisition to strongly constrain the problem, thereby ensuring smooth and robust motion fields. The main contribution of this paper is to formally associate the C-arm angulation from which the heart is observed to the expected parametric motion field. Once this learning phase performed, every new sequence is associated to a fixed motion model, known up to a scale and an origin. The quality of the extracted velocities is visually evaluated over a large representative database. We also present an application of our approach to the motion-based classification of an angiographic exam into right or left artery tree.

Large and performant neural networks are often overparameterized and can be drastically reduced in size and complexity thanks to pruning. Pruning is a group of methods, which seeks to remove redundant or unnecessary weights or groups of weights in a network. These techniques allow the creation of lightweight networks, which are particularly critical in embedded or mobile applications. In this paper, we devise an alternative pruning method that allows extracting effective subnetworks from larger untrained ones. Our method is stochastic and extracts subnetworks by exploring different topologies which are sampled using Gumbel Softmax. The latter is also used to train probability distributions which measure the relevance of weights in the sampled topologies. The resulting subnetworks are further enhanced using a highly efficient rescaling mechanism that reduces training time and improves performance. Extensive experiments conducted on CIFAR show the outperformance of our subnetwork extraction method against the related work.

Human brain connection map is far from being complete. In particular the study of the superficial white matter (SWM) is an unachieved task. Its description is essential for the understanding of human brain function and the study of the pathogenesis associated to it. In this work we developed a method for the automatic creation of a SWM bundle multi-subject atlas. The atlas generation method is based on a cortical parcellation for the extraction of fibers connecting two different gyri. Then, an intra-subject fiber clustering is applied, in order to divide each bundle into sub-bundles with similar shape. After that, a two-step inter-subject fiber clustering is used in order to find the correspondence between the sub-bundles across the subjects, fuse similar clusters and discard the outliers. The method was applied to 40 subjects of a high quality HARDI database, focused on the left hemisphere fronto-parietal and insula brain regions. We obtained an atlas composed of 44 bundles connecting 22 pair of ROIs. Then the atlas was used to automatically segment 39 new subjects from the database.