Segmentation of the aorta in 4D flow MRI using 4D convolutional kernels and learning from sparse annotations

SummaryAI generated

Researchers developed an automatic 4D U-Net model based on a parameter-efficient hybrid 4D convolutional kernel for the segmentation of the ascending, arch, and proximal descending parts of the aorta in medical images. This method innovatively uses sparse 4D labels derived from existing 2D expert contours and centerlines for training, avoiding the need for dense annotation. Experimental data included 268 scans from 8 medical centers and 2 manufacturers, and the model was validated on an internal test set and external post-contrast enhancement datasets. Evaluation results showed that the model outperformed frame-based 3D networks, static PC-MRA, and registration-based propagation methods in terms of Dice score (0.927 internally, 0.911 externally) and consistency index with expert contours (ICC ≥ 0.954/0.980). The model achieved reproducible time-resolved aortic segmentation across centers, manufacturers, and independent post-contrast enhancement data, and the related results have been published.

Related eventsRELATED EVENTS

2026-09-09 05:29OpenAI claims huge maths breakthrough on a famed ‘Millennium Problem’
2026-09-09 06:11OpenAI claims to have solved the Navier-Stokes equation problem, but its originality is questioned by the academic community
2026-09-08 22:00Google DeepMind Maps 9 Billion Possible DNA Variants

Key entitiesKEY ENTITIES

arXiv

SignalsSIGNALS

Keyword heat

arXiv1

All reports (1)SOURCES

↗

A arXiv cs.CV en 2026-09-07 12:00

The researchers developed an automatic 4D U-Net based on parameter-efficient hybrid 4D convolutional kernels for segmenting the ascending, arch, and proximal descending parts of the aorta. This method utilizes sparse 4D labels derived from existing 2D expert contours and centerlines, eliminating the need for dense 4D annotation. The training data consisted of 268 scans from 8 centers and 2 manufacturers, and the model was evaluated on an internal test set (32 scans) and an external post-contrast enhancement dataset (30 scans). Compared with time-resolved annotation, the model achieved Dice scores of 0.927 and 0.911 in internal and external tests, respectively, outperforming frame-based 3D networks, static PC-MRA, and registration-based propagation methods; indicators such as peak velocity, net flow, shear stress, and diameter showed excellent consistency with expert contours (ICC ≥ 0.954/0.980). The model achieved reproducible time-resolved aortic segmentation across centers, manufacturers, and independent post-contrast enhancement data, and has been published on public platforms.