All research

Biomedical devices & simulation · 2025

Rapid wall shear stress prediction for abdominal aortic aneurysms

A deep learning surrogate that predicts time-averaged wall shear stress on abdominal aortic aneurysms from 3D vessel geometry.

Qatar University IRCC-2020-002 · Qatar National Research Fund NPRP13S-0108-200024

A three-dimensional aorta coloured by time-averaged wall shear stress is rotated around its vertical axis in 30-degree steps, producing twelve images from 0 to 330 degrees.
Figure 4 from Faisal et al. (2025), cropped from the article page. A 3D aortic geometry is converted into 12 rendered views.View full-size figureSource articleCC BY 4.0

Approach

Can deep learning provide a fast alternative to CFD for mapping wall shear stress on aortic aneurysms?

Low wall shear stress is linked with aneurysm progression and rupture risk, but computational fluid dynamics (CFD) is too time-consuming for rapid clinical use. This study built patient and synthetic aneurysm geometries, generated CFD references, converted each 3D shape into multiple 2D views, and trained MultiViewUNet to predict the wall shear stress pattern.

  1. Build geometries and run CFD
  2. Render multi-view images
  3. Predict wall shear stress

Building the aneurysm dataset and CFD reference

Medical & Biological Engineering & Computing · 2025

CT scans from 23 patients with abdominal aortic aneurysm, provided by Hamad General Hospital, were segmented into 3D aortic geometries. With guidance from clinicians, 10 synthetic variants were created from each patient geometry, giving 253 geometries in total.

Transient CFD simulations in ANSYS CFX covered three cardiac cycles, and the third cycle was used to calculate time-averaged wall shear stress (TAWSS). Each synthetic geometry took about one hour, and the synthetic set required 230 hours and 8 TB of storage.

Aortic geometries
253
Patient geometries, all held for testing
23
CFD time per synthetic geometry
~1 h
Training and test split described in Section 2.8 of Faisal et al. (2025).
TrainingTestingTotal
Patient geometries02323
Synthetic geometries19040230
All geometries19063253

Turning 3D geometry into image pairs

Medical & Biological Engineering & Computing · 2025

Each geometry was rendered from 12 angles by rotating it around its vertical axis in 30-degree steps, and the images were resized to 256 × 256 pixels. This turned a geometry-to-geometry problem into image-to-image translation and removed the need for every aorta to have the same number of mesh nodes.

Surface curvature was added to the input as a colour map because it describes aneurysm shape directly. Compared with shaded geometry alone, curvature reduced the normalised absolute error and raised the effective peak signal-to-noise ratio by more than 6 dB.

Rendered views per geometry
12
Rotation between views
30°
Image size used by the model
256 × 256
Results for the two input representations, from Section 3 of Faisal et al. (2025).
NMAENMSENRMSEEPSNR
Shaded geometry0.402%0.086%1.316%23.67 dB
Geometry with curvature0.362%0.021%0.643%29.79 dB
A plain 3D aorta passes through a curvature calculation and colour mapping step, producing the same aorta coloured by surface curvature.
Figure 3 from Faisal et al. (2025), cropped from the article page. Surface curvature is added to the raw geometry as a colour map.View full-size figureSource articleCC BY 4.0

Predicting wall shear stress with MultiViewUNet

Medical & Biological Engineering & Computing · 2025

MultiViewUNet is a five-level image-to-image network that combines full-scale connections between its encoder and decoder with additional connections inside each path. It was trained with a loss calculated only over pixels belonging to the aorta, so the white background did not affect learning.

On the 63 held-out geometries, it gave the lowest image-based errors among the compared models. Predictions closely followed CFD on patient and synthetic geometries, with the largest differences in regions of disturbed flow.

Normalised mean absolute error
0.362%
Effective peak signal-to-noise ratio
29.79 dB
Inference time on an NVIDIA T4 GPU
11.85 ms
Model comparison from Table 1 of Faisal et al. (2025). PointNet metrics were calculated directly from 3D geometry, so they are not fully comparable with the image-based models.
NMAENMSENRMSEEPSNRInference time
U-Net0.414%0.086%1.305%23.67 dB8.02 ms
Pix2Pix0.379%0.032%0.805%27.3 dB8.13 ms
PointNet4.52%0.573%7.57%18.3 dB6.45 ms
MultiViewUNet0.362%0.021%0.643%29.79 dB11.85 ms
MultiViewUNet diagram with five encoder levels and four decoder levels joined by dense curved skip connections inside and across the encoder and decoder paths.
Figure 6 from Faisal et al. (2025), cropped from the article page. The MultiViewUNet architecture.View full-size figureSource articleCC BY 4.0
Seven patient aortic geometries shown twice. CFD wall shear stress maps in the top row closely resemble the model predictions in the bottom row.
Figure 9 from Faisal et al. (2025), cropped from the article page. CFD reference (top) and predicted wall shear stress (bottom) on patient geometries.View full-size figureSource articleCC BY 4.0
Seven synthetic aortic geometries shown twice. CFD wall shear stress maps in the top row closely resemble the model predictions in the bottom row.
Figure 10 from Faisal et al. (2025), cropped from the article page. CFD reference (top) and predicted wall shear stress (bottom) on synthetic geometries.View full-size figureSource articleCC BY 4.0
Seven aortic geometries comparing CFD and predicted wall shear stress, with red arrows marking regions where the prediction differs from the reference.
Figure 11 from Faisal et al. (2025), cropped from the article page. Less accurate cases, with differences marked by red arrows.View full-size figureSource articleCC BY 4.0

My contribution

I helped implement and integrate the deep learning model, evaluated successive versions, and troubleshot the modelling workflow with the core team.

Study scope

CFD was the reference, so agreement depends on the geometries, boundary conditions and simulation assumptions used. The dataset was relatively small, and errors were larger in disturbed flow. The model outputs 256 × 256 images rather than a reconstructed 3D surface, and other haemodynamic measures remain future work.