3D segmentation of perivascular spaces on T1-weighted 3 Tesla MR images with convolutional autoencoder and U-shaped neural networks
Keywords: perivascular space, deep learning, U-net, MRI, brain cohort, segmentation
TL;DR: We implemented a deep learning model for the 3D segmentation of perivascular space in deep white matter
Abstract: For its involvement in cognitive deterioration and dementia, assessment of enlarged perivascular spaces (PVSs) has become a major area of interest. We implemented a deep learning model for the 3D segmentation of PVSs in deep white matter. It was trained and tested using T1-weighted magnetic resonance imaging data from 1,832 young adults. The model was trained first based on a CNN autoencoder with the full dataset then with a U-net like architecture trained with a subset of 40 T1-weighted MRI manually annotated images. The Dice coefficient (from a separate test subset of 10 images) was 0.64 for cluster detection. Dice values above 0.90 were reached for detecting PVSs larger than 10 mm3. Using the full dataset, the predicted PVS load showed a high degree of agreement with a semi-quantitative visual rating. Finally, we demonstrated the interoperability of this model using a second dataset.
Paper Type: validation/application paper
Primary Subject Area: Segmentation
Secondary Subject Area: Learning with Noisy Labels and Limited Data
Paper Status: based on accepted/submitted journal paper
Source Code Url: The code will be publish at the time of the conference
Data Set Url: Access to MRI-share dataset are to be asked to www.i-share.fr
BIL&GIN dataset is open for collaboration upon request (https://www.gin.cnrs.fr/en/current-research/axis2/bilgin-en/)
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
4 Replies
Loading