The availability of a high-performance tool for segmenting multiple sclerosis lesions provides radiologists with reliable and reproducible metrics, leading to improved patient care. To make this type of tool more accessible in clinical settings, we have proposed lightweight and high-performance convolutional neural network architectures capable of learning on affordable workstations with a limited number of training examples, in a short amount of time, while minimizing the risk of overfitting. We implemented techniques aimed at minimizing the need for training data through self-learning and semi-supervised learning, while taking data quality into account, and found that, ultimately, very few annotated scans were sufficient. We also present a method to increase the detection of small lesions, which are more difficult to segment and more likely to be missed by the radiologist. This thesis is part of a research effort to maximize the benefits of automatic lesion segmentation for radiologists, with the goal of promoting the adoption of such tools in clinical practice.
Author : Alexandre Fenneteau
Management team :
- D.Helbert
- P.Bourdon
- M.Naudin
- C.Guillevin
- C.Fernandez
- C.Habbas
- R.Guillevin
Thesis defended on November 25, 2022
Accepted publications:
- A.Fenneteau, P.Bourdon, D. Helbert, C.Fernandez-Maloigne, C. Habas, “Learning a CNN on multiple sclerosis lesion segmentation with self-supervision” in 3D Measurement and Data Processing session, IS&T Electronic Imaging 2020 Symposium, , San Francisco, US, Jan 2020. Best paper award
- Alexandre Fenneteau, Pascal Bourdon, David Helbert, Christine Fernandez-Maloigne, Christophe Habas, Rémy Guillevin, « Investigating efficient CNN architecture for multiple sclerosis lesions segmentation » , Journal of Medical imaging, 014504 2021 (IF 1.58) DOI : https://doi.org/10.1117/1.JMI.8.1.014504