Portal de Periódicos da CAPES

OP04.06: Improving and accelerating the assembly of fetal brain MRI databases

2023; Wiley; Volume: 62; Issue: S1 Linguagem: Inglês

10.1002/uog.26481

1469-0705

Clément Godard, G. Perdu, R. Cremese, Thomas Blanc, L. Bobet, Gérome C. Gauchard, D. Grévent, Bassam Hajj, L. J. Salomon, Jean‐Baptiste Masson,

Domain Adaptation and Few-Shot Learning

Fetal MRI has emerged as a valuable tool for investigating neurological development in fetuses with congenital disorders. However, automated segmentation of the developing brain are necessary to improve diagnostics. To achieve this goal, we need to enhance and expedite fetal brain annotation, which will improve automated machine learning pipelines designed to perform tissue segmentation. We present DIVA-Hesperos, a software platform designed to enhance and expedite medical image annotations and automated segmentation algorithms. The platform consists of a modified version of DIVA, a software for visualising 3D data in virtual reality (VR), along with a tablet application called Hesperos. This app enables retrieval of DIVA annotations, data navigation under radiologist-friendly interfaces, and annotations with various assistance tools. Hesperos-DIVA also allows for fusion of segmentations with raw images and the placement of landmarks to indicate anomaly areas. We demonstrate our approach on the 80 brains of the FETA brain tissue database. We show that we can improve the annotation of brain tissues by more faithfully respecting their 3D structures and also improve automatic segmentation pipelines performance. Finally, we show transfer learning between the FETA database and 25 stacks taken at the Lumiere Foundation (ClinicalTrials.gov NCT01092949). By combining virtual reality, tablet apps, and self-supervised learning, it is possible to create accelerated approaches to the automatic processing of fetal MRI images. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.