Approche semi-supervisée pour l'analyse automatique de la moelle épinière sur imagerie par résonance magnétique

Abstract

This study uses 2D convolutional neural networks to automate spinal cord segmentation from T2-weighted MRIs. The database consists of 106 sagittal MRI scans from 94 patients with traumatic spinal cord injuries. Two complete methods integrating data preprocessing and automatic segmentation have been developed. One of these methods aims for the best possible segmentation of the spinal cord from a clinical point of view. This method uses an innovative approach where the MRIs are analyzed in series under the axial and sagittal plane by two different networks. The results are compared with those of Deepseg 2D from the Spinal Cord Toolbox, which is a method representing the state of the art. Our method achieves significantly better results than Deepseg 2D, with an average Dice coefficient of 0.95 against 0.88 for Deepseg 2D (P < 0.001). The other approach aims to develop and evaluate a semi-supervised learning method. In this approach, two networks are trained in parallel, where each network uses a binarized version of the segmentations produced by the other network to improve its own training. To evaluate this approach, four pairs of networks are trained with 25%, 50%, 75% and 100% of the annotated data respectively, then the results are compared with those of networks using only supervised learning. The results show a considerable improvement over the supervised baseline, an improvement that is certainly more marked when a lesser portion of the data is annotated, but which nevertheless remains with 100% of the annotated data. Two approaches have therefore been successfully developed to segment the human spinal cord, one presenting a supervised technique which achieves segmentation results exceeding the state of the art, and the other presenting a semi-supervised technique which performs well with variable amounts of data.

Date	27 Jun 2022
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Luc Duong (Supervisor) & Jean Marc Mac-Thiong (Co-supervisor)