Robust Segmentation of Corpus Callosum in Multi-Scanner pediatric T1-w MRI using Transfer Learning

The Corpus Callosum (CC) has been intensively studied in adults and it has been shown, for instance, that there is a relationship between its morphology and some neurological diseases. Fewer morphological studies have been conducted on children showing that variations in size and shape might be correlated with Multiple Sclerosis. T1-weighted MR scans are usually employed since the CC is entirely visible and distinguishable in the Mid-Sagittal Plane (MSP). State-of-the-art segmentation algorithms for adults, such as CCSeg and ART-yuki, are difficult to use with pediatric images due to the low contrast-to-noise ratio, short acquisition time (low resolution), presence of other parts of the body. In this work we present a robust method to segment the Corpus Callosum in Magnetic Resonance Images (MRI) based on Convolutional Neural Networks (CNN) and Transfer Learning. The thesis was carried out throughout an internship of six months at LTCI (Laboratoire Traitement et Communication de l’Information) of the Ecole Nationale Supérieure des Télécommunications (TELECOM ParisTech), and is part of a much larger joint French-Brazilian project, named ANR-FAPESP Project STAP, which main topic is spatio-temporal analysis of pediatric MRI. We first train a CNN on adult T1-w MR images, then perform transfer learning on pediatric images with different scanners and protocols to improve generalization power. Pediatric radiologists of the Bicêtre Hospital of Paris, who have collaborated in this work, suppose that the Corpus Callosum decreases in size and shape over time in Children who have MS, rather than in those affected by other inflammatory diseases, such as ADEM (Acute Disseminated Encephalomyelitis), transverse myelitis and optic myelitis. The second step of our research work is to verify these assumptions, thanks to the results obtained by our proposed method