A MedNeXt-Based Deep Learning Model for Automated Segmentation of Perivascular Spaces

0 2 minutes read

A MedNeXt Based Deep Learning Model for Automated Segmentation of Perivascular.png

[Submitted on 27 Aug 2025]

Authors:Zhen Xuen Brandon Low, Rory Zhang, Hang Min, William Pham, Lucy Vivash, Jasmine Moses, Miranda Lynch, Karina Dorfman, Cassandra Marota, Shaun Koh, Jacob Bunyamin, Ella ROWTHORN, Alexa Jarema, Himashi Peiris, Sharon L. Naismith, TERENCE J. O’Brien, Ying XIA, Ming law, Benjamin Sinclair

View the PDF file for the paper titled Megnet-PVS: a deep-based educational model for retail mechanism for the spaces surrounding the vessels, written by Zhen Xuen Brandon Low and 22 other authors

PDF view

a summary:The areas surrounding the blood vessels (PVS) are increasingly recognized as the vital indicators of the disease of the small cerebral bowl, Alzheimer’s disease, stroke, and nervous transmission associated with aging. However, PVS manual retail takes a long time and is subject to a moderate adjective between the center, while the current deep automated learning models have moderate performance and usually fail to generalize through clinical magnetic and research photography collections. We have adapted the mednext-l-k5, a three-dimensional recovery network inspired by transformers, in order to divide the automatic PVS. Two models were trained: one uses a homogeneous data collection of 200 T2 Port (T2W) MRI from the MRI data collection (T1W) that increases seven studies through scan -scanners. Typical performance was evaluated using 5 -fold internal health verification (5FCV) and validation of the vacation in one location (Losocv). Mednext-L-K5 models trained on T2W pictures of the HCP Vouxel Vouxel Blossom Data Set from 0.88 +/- 0.06 (White Article, WM), similar to a reliability between the amount between these data, and the highest reported in the literature. The same models trained on T1W pictures of the HCP Aging Data set have achieved much lower than 0.58 +/- 0.09 (WM). Under Losocv, the form of blossom scores at 0.38 +/– 0.16 (WM) and 0.35 +/- 0.12 (BG), dozens of bloc at the level of the mass from 0.61 +/- 0.19 (WM) and 0.62 +/– 0.21 (BG). Mednext-L-K5 provides an effective solution to the MRI T1W and T2W data collections. Mednext-L-K5 did not outperform NNU-Net, indicating that the attention-based mechanisms in the models inspired by transformers to provide a global context are not required for high accuracy in PVS segmentation.