Évaluation volumétrique de lésions gliales en imagerie par résonance magnétique 3 tesla : segmentation manuelle et segmentation automatique par algorithme d'intelligence artificielle
Year: 2023
Type: dissertation
Abstract: Introduction: MRI plays a central role in the diagnosis and monitoring of glial lesions. volumetry of brain lesions can be time-consuming and tedious for radiologists in their daily practice. In this study, the aim was to develop a TR-UNet type artificial intelligence capable of performing automatic volumetry both in the initial assessment and in the follow-up of glial lesions, regardless of the patient's grade and operative status. Materials and Methods: Fifty patients with glial lesions were manually segmented by a radiologist to serve as a training dataset for artificial intelligence, while seven patients with glioblastoma, previously segmented by an engineer, served as a test dataset. The Dice score was the primary metric for evaluating automatic volumetry. Results: The average Dice score in the test population ranged from 0,83, indicating satisfactory automatic segmentations. Discrepancies varied, primarily depending on the appearance and heterogeneity of the lesions. It was shown that the TR-UNet artificial intelligence model outperformed a conventional UNet artificial intelligence model in glial lesion segmentation under real-world conditions. Conclusion: TR-UNet-type artificial intelligence, after targeted training, is capable of automatically segmenting glial lesions with relatively good accuracy, depending primarily on the appearance and homogeneity of the target lesions. In routine practice, an artificial intelligence solution appears to be a suitable alternative to manual segmentation, which is far too time-consuming to be viable in real clinical settings.
Source: HAL
Author: Millet, Léo
Institution: Université de Poitiers
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7018806594
