PrePostEVAR_CTA_Annotation

This repository contains trained 3D nnU-Net models for automated analysis of pre- and post-endovascular aneurysm repair (EVAR) computed tomography angiography (CTA).

Table of Contents

• Model Details
• Uses
• Bias, Risks, and Limitations
• Training Details
• Evaluation
• Environmental Impact
• Citation

Model Details

Model Description

The models weights provided are used in a 3D U-Net with the automatic dataset-based configuration provided by nnU-Net framework. Four independent tasks are provided, covering two main categories:

Anatomical segmentation Segmentation-based landmark localization

Separate models were trained for pre-EVAR and post-EVAR CTA because these imaging scenarios differ substantially in anatomy, appearance, and target structures.

Each task was trained using five-fold cross-validation, resulting in five independently trained models per task and 20 trained models in total. The repository contains the checkpoint_best.pth weights corresponding to the best validation performance for each fold.

Repository folder	Task	Input	Output
Dataset100_SegmentationPreEVAR	Anatomical segmentation	Pre-EVAR CTA	Aortic lumen and intraluminal thrombus
Dataset101_LandmarksPreEVAR	Landmark localization	Pre-EVAR CTA	Abdominal aorta + 6 anatomical landmarks
Dataset200_SegmentationPostEVAR	Anatomical segmentation	Post-EVAR CTA	Aortic lumen, excluded aorta and endoleak
Dataset201_LandmarksPostEVAR	Landmark localization	Post-EVAR CTA	Abdominal aorta + 9 anatomical/endograft-related landmarks

Each model directory contains five fold-specific weights:

DatasetXXX_TaskName/
└── nnUNetTrainer__nnUNetPlans__3d_fullres/
    ├── dataset.json
    ├── plans.json
    ├── fold_0/
    │   └── checkpoint_best.pth
    ├── fold_1/
    │   └── checkpoint_best.pth
    ├── fold_2/
    │   └── checkpoint_best.pth
    ├── fold_3/
    │   └── checkpoint_best.pth
    └── fold_4/
        └── checkpoint_best.pth

Anatomical segmentation

Pre-EVAR segmentation

Model folder: Dataset100_SegmentationPreEVAR

The pre-EVAR models were trained to segment:

Label	Structure
0	Background
1	Aortic lumen
2	Intraluminal thrombus

Voxel-wise predictions were converted into discrete segmentation masks by assigning each voxel to the class with the highest predicted probability.

Configuration

Parameter	Value
Target spacing	0.79 × 0.70 × 0.79 mm
Median image size	512 × 640 × 512 voxels
Patch size	128 × 160 × 112 voxels

Post-EVAR segmentation

Model folder: Dataset200_SegmentationPostEVAR

The post-EVAR models were trained to segment:

Label	Structure
0	Background
1	Aortic lumen
2	Excluded aorta
3	Endoleak (EL)

The excluded aorta was defined as the aortic volume outside the stent-graft lumen, excluding the stent graft itself.

Voxel-wise predictions were converted into discrete segmentation masks by assigning each voxel to the class with the highest predicted probability.

Endoleak detection

Endoleak detection was derived directly from the post-EVAR anatomical segmentation models.

Unlike the other anatomical structures, the endoleak output was retained as a voxel-wise probability map. Binary endoleak masks were generated using the following threshold:

P(EL) ≥ 0.5

After thresholding, three-dimensional connected components containing fewer than 10 voxels were removed to reduce isolated false-positive predictions.

Configuration

Parameter	Value
Target spacing	0.77 × 0.70 × 0.77 mm
Median image size	512 × 591 × 512 voxels
Patch size	128 × 128 × 128 voxels

---

Landmark localization

Landmark localization was formulated as a segmentation task rather than direct coordinate regression.

Each manually annotated landmark coordinate was converted into a small cubic binary mask with an edge length of approximately 3 mm.

Because the masks were defined in voxel space and CTA voxel spacing varied between scans, their exact physical dimensions could vary slightly after rounding to the nearest voxel.

This formulation enabled the use of the standard nnU-Net segmentation pipeline while reducing the extreme class imbalance that would result from representing each landmark as a single voxel.

After prediction:

1. Voxel-wise predictions were converted into discrete masks by assigning each voxel to the class with the highest predicted probability.
2. The final coordinate of each landmark was calculated as the centroid of the predicted mask corresponding to that landmark class.

Pre-EVAR landmark localization

Model folder: Dataset101_LandmarksPreEVAR

The model predicts an abdominal aorta class together with six anatomical landmark classes:

Label	Structure / landmark
0	Background
1	Abdominal aorta
2	Celiac trunk
3	Superior mesenteric artery
4	Renal arteries
5	Start of the abdominal aortic aneurysm (AAA)
6	Aortic bifurcation
7	End of the AAA

Configuration

Parameter	Value
Target spacing	0.79 × 0.70 × 0.79 mm
Median image size	145 × 372 × 154 voxels
Patch size	96 × 256 × 96 voxels

Post-EVAR landmark localization

Model folder: Dataset201_LandmarksPostEVAR

The model predicts an abdominal aorta class together with nine anatomical and endograft-related landmark classes:

Label	Structure / landmark
0	Background
1	Abdominal aorta
2	Celiac trunk
3	Superior mesenteric artery
4	Renal arteries
5	Start of the abdominal aortic aneurysm (AAA)
6	Aortic bifurcation
7	End of the AAA
8	Proximal prosthesis landing
9	Proximal prosthesis fabric component
10	Distal prosthesis landing

Configuration

Parameter	Value
Target spacing	0.77 × 0.70 × 0.77 mm
Median image size	142 × 364 × 158 voxels
Patch size	96 × 256 × 96 voxels

---

Dataset fingerprints

The main task-specific configurations automatically determined by nnU-Net are summarized below.

Parameter	Pre-EVAR Segmentation	Pre-EVAR Landmarks	Post-EVAR Segmentation	Post-EVAR Landmarks
Dataset	Dataset100_SegmentationPreEVAR	Dataset101_LandmarksPreEVAR	Dataset200_SegmentationPostEVAR	Dataset201_LandmarksPostEVAR
Target spacing (mm)	0.79 × 0.70 × 0.79	0.79 × 0.70 × 0.79	0.77 × 0.70 × 0.77	0.77 × 0.70 × 0.77
Median image size (voxels)	512 × 640 × 512	145 × 372 × 154	512 × 591 × 512	142 × 364 × 158
Patch size (voxels)	128 × 160 × 112	96 × 256 × 96	128 × 128 × 128	96 × 256 × 96
Foreground classes	2	7	3	10

• Developed by:
  Cardiovascular Image Unit | Cardiovascular diseases research team at Vall d'Hebron Institut de Recerca (VHIR)

• Funded by:
  This study was funded by the "Red de Investigación Vascular" of the "Sociedad Española de Angiología y Cirugía Vascular", the Instituto de Salud Carlos III (CP24/00121, co-funded by the European Union), the Spanish Ministry of Science, Innovation and Universities (FORT23/00034) and the Biomedical Research Networking Center on Cardiovascular Diseases (CIBERCV). Juan Garrido-Oliver has received funding from Secretaria d'Universitats i Recerca del Departament de Recerca i Universitats de la Generalitat de Catalunya i del Fons Europeu Social Plus (AGAUR-FI 2023 FI-1 00322 Joan Oró).

• Model type:
  Deep Learning model (CNN)

• License:
  cc-by-nc-sa-4.0

Uses

Direct Use

Follow the official instructions to install the nnU-Net framework

• Once nnU-Net is installed, make sure you have created the nnUNet_results directory and assigned it to the corresponding environment variable as required by nnU-Net.

Then, copy the folder DatasetXXX_YYYYY into the nnUNet_results directory.

This folder contains the trained model weights for abdominal aorta segmentation and landmarks annotation in pre- and post-EVAR contrast-enhanced computed tomography angiography (CTA) images.

For inference, you need to specify a folder with the images in the proper nnU-Net format (see in their guide).

CUDA_VISIBLE_DEVICES=0 nnUNetv2_predict -i INPUT_FOLDER -o OUTPUT_FOLDER -d XXX -c 3d_fullres -chk checkpoint_best.pth

Remember, XXX defines the models you are using (i.e. 100 for pre-EVAR segmentation, 101 for pre-EVAR landmark annotation, 200 for post-EVAR segmentation, and 201 for post-EVAR landmark annotation)

Out-of-Scope Use

Use it with CTA images that include the abdominal aorta. It is not recommended to use it with whole-body images (unless your GPU and RAM are large) or other modalities.

Bias, Risks, and Limitations

This model was trained using images from patients with abdominal aortic aneurysms indicated for surgery or post-EVAR CTAs. No healthy subjects were included in training, therefore there might be a bias towards better segmentation in patients with aneurysms.

Training Details

One hundred and thirty-five pre-EVAR CTAs and 160 post-EVAR CTAs from 135 patients from one center were used in training.

Network architecture

All models follow the general 3D U-Net architecture automatically configured by nnU-Net, with task-specific parameters adapted to the corresponding dataset fingerprint.

Each resolution stage contains two convolutional blocks composed of:

• 3 × 3 × 3 convolutions
• Instance normalization
• Leaky ReLU activation

The encoder uses strided convolutions for downsampling across six resolution stages, while the decoder uses five transposed-convolution upsampling stages to progressively recover the spatial resolution.

Training configuration

All four tasks were trained independently using the nnU-Net training framework.

Parameter	Value
Architecture	3D U-Net
Optimizer	Stochastic gradient descent (SGD)
Momentum	Nesterov momentum = 0.99
Loss function	Dice + Cross-entropy
Batch size	2 patches
Initial learning rate	0.01
Learning-rate schedule	Polynomial decay
Training length	1000 epochs
Iterations per epoch	250
Cross-validation	5 folds

Evaluation

Evaluation is described in the research article Garrido-Oliver et al., 2026, Eur J Vasc Endovasc Surg (see references).

Testing Data, Factors & Metrics

Aortic segmentations and landmarks were evaluated using the cross-validation data from one center and an external validation dataset from another center that included data from 30 patients including 30 pre-EVAR CTAs and 40 post-EVAR CTAs.

Testing Data

Thirty CTA images were annotated by two different endovascular surgeons using 3DSlicer.

Metrics

Segmentation performance was evaluated using Dice score (DS), an overlap volume-based metric, and Hausdorff distance 95th percentile (HD), a surface distance-based metric. Automatic segmentations were compared with each observer annotation and a further test was designed to identify whether the availability of an DL segmentation as a starting point improved the inter-observer agreement of manual segmentation.

Landmark annotation was evaluated using the Euclidean distance between manual annotations and model's predictions. Also, the absolute axial distance and the signed axial distance were computed to identify bias toward craneo-caudal direction.

Results

Table 1. Results of manual interobserver variability and DL accuracy in segmentation for the center 1 and center 2 CTAs.
* Significant difference (p<0.05) between interobserver variability and DL errors.
† Significant difference (p<0.05) between DL error in C1 and DL error in C2 data.
DS: Dice score. HD95: 95th percentile of the Hausdorff distance in mm. C1: Training set (cross-validation). C2: External validation set.

Structure	Interobserver variability DS	Interobserver variability HD95	DL vs. manual C1 DS	DL vs. manual C1 HD95	DL vs. manual C2 DS	DL vs. manual C2 HD95
Pre-EVAR CTA (N)	30		135		30
Aortic lumen	0.955 [0.935, 0.970]	1.89 [1.76, 2.07]	0.969 [0.952, 0.978]*	1.73 [1.58, 1.98]	0.955 [0.938, 0.967]†	1.93 [1.74, 2.19]†
Intraluminal thrombus	0.902 [0.861, 0.931]	3.10 [2.63, 8.32]	0.922 [0.888, 0.947]	2.62 [1.99, 3.97]*	0.907 [0.879, 0.934]	3.02 [2.42, 4.23]
Whole aorta	0.950 [0.944, 0.967]	2.47 [2.21, 2.97]	0.962 [0.948, 0.974]	2.34 [2.02, 2.83]	0.953 [0.939, 0.962]†	2.40 [2.10, 2.93]
Post-EVAR CTA (N)	30		160		40
Aortic lumen	0.879 [0.868, 0.933]	2.12 [1.82, 3.02]	0.935 [0.908, 0.954]*	1.73 [1.58, 2.25]*	0.914 [0.875, 0.930]†	2.14 [1.79, 3.06]
Excluded aorta	0.926 [0.913, 0.934]	3.36 [2.40, 4.43]	0.940 [0.916, 0.953]	2.58 [2.17, 3.70]*	0.934 [0.911, 0.943]	2.64 [2.39, 3.66]
Whole aorta	0.918 [0.890, 0.932]	4.11 [3.21, 5.51]	0.938 [0.912, 0.950]*	3.09 [2.48, 15.13]	0.923 [0.905, 0.940]†	3.38 [2.88, 4.99]

Table 2. Euclidean distance (mm), represented as median [1st-3rd] quartiles, between landmarks annotated by two observers, and DL vs manual reference in both centers.
* Significant difference (p<0.05) between interobserver difference and DL errors. No significant differences were found in DL error between C1 and C2 data.
SMA: superior mesenteric artery; AAA: abdominal aortic aneurysm. C1: Training set (cross-validation). C2: External validation set.

Landmark	Pre-EVAR Interobserver variability	Pre-EVAR DL vs manual C1	Pre-EVAR DL vs manual C2	Post-EVAR Interobserver variability	Post-EVAR DL vs manual C1	Post-EVAR DL vs manual C2
Celiac trunk	3.35 [2.03, 4.14]	2.38 [1.88, 3.29]*	2.01 [1.63, 3.90]	2.43 [2.02, 3.97]	2.52 [1.76, 3.54]	3.57 [1.73, 11.18]
SMA	2.87 [2.23, 3.96]	2.94 [2.13, 4.10]	2.80 [2.06, 4.38]	3.05 [2.48, 4.00]	2.89 [2.00, 4.14]	2.88 [1.72, 4.07]
Renal arteries	2.03 [1.35, 3.01]	2.44 [1.88, 3.33]	2.55 [2.26, 3.45]	1.93 [1.45, 2.97]	2.49 [1.97, 4.22]*	2.84 [1.84, 5.52]*
Proximal origin AAA	7.13 [4.76, 10.90]	5.82 [3.98, 9.15]	4.53 [3.55, 6.15]*	5.37 [3.78, 10.82]	4.90 [2.84, 8.22]	4.54 [3.19, 5.73]
Aortic bifurcation	3.89 [2.25, 6.66]	5.90 [4.11, 7.60]*	6.74 [5.37, 8.27]*	3.79 [2.27, 5.75]	6.98 [5.38, 8.85]*	7.33 [5.97, 8.94]*
Distal end AAA	9.13 [4.56, 13.03]	9.64 [4.17, 24.31]	12.78 [7.27, 30.32]	8.33 [5.77, 14.14]	10.20 [5.11, 23.89]	8.33 [3.74, 12.86]
Prox. land. EVAR	-	-	-	6.18 [4.33, 8.84]	4.25 [2.83, 6.08]*	3.70 [2.55, 7.11]*
Fabric EVAR	-	-	-	4.01 [2.57, 6.06]	4.27 [2.81, 6.98]	4.31 [2.32, 14.87]
Distal land. EVAR	-	-	-	4.04 [2.49, 6.64]	11.21 [3.68, 28.37]*	14.27 [3.76, 24.84]*
All landmarks	3.61 [2.19, 6.74]	3.69 [2.31, 6.87]	4.21 [2.33, 6.70]	3.87 [2.34, 7.11]	4.26 [2.48, 7.80]	4.34 [2.31, 9.41]

Table 3. Absolute distance in the axial direction (mm), represented as median [1st-3rd] quartiles, between landmarks annotated by two observers, and DL vs manual reference in both centers.
* Significant difference (p<0.05) between interobserver difference and DL errors.
† Significant difference (p<0.05) between DL error in C1 and DL error in C2 data.
SMA: superior mesenteric artery; AAA: abdominal aortic aneurysm. C1: Training set (cross-validation). C2: External validation set.

Landmark	Pre-EVAR Interobserver variability	Pre-EVAR DL vs. manual C1	Pre-EVAR DL vs. manual C2	Post-EVAR Interobserver variability	Post-EVAR DL vs. manual C1	Post-EVAR DL vs. manual C2
Celiac trunk	1.41 [0.70, 2.50]	1.41 [0.69, 2.31]	1.75 [1.26, 3.57]	0.75 [0.26, 1.73]	1.28 [0.43, 2.74]†	2.62 [0.83, 8.65]*†
SMA	1.45 [0.79, 2.76]	1.76 [0.68, 2.85]	1.81 [1.12, 2.86]	1.25 [0.54, 2.20]	1.64 [0.70, 2.99]	1.94 [0.92, 3.69]*
Renal arteries	0.89 [0.36, 1.58]	1.68 [0.85, 2.46]*	1.80 [0.82, 2.69]*	0.87 [0.54, 1.78]	1.88 [0.85, 3.20]*	2.31 [0.96, 4.19]*
Proximal origin AAA	2.80 [1.86, 7.08]	3.63 [1.91, 7.04]	3.79 [2.24, 5.18]	2.95 [1.76, 5.77]	3.10 [1.38, 6.65]	2.60 [1.44, 5.08]
Aortic bifurcation	1.91 [1.36, 2.91]	2.83 [1.10, 4.15]	4.05 [2.01, 6.67]*†	2.59 [1.30, 4.33]	2.74 [1.21, 4.69]	3.41 [1.61, 5.36]
Distal end AAA	4.90 [2.28, 9.49]	6.17 [2.21, 13.51]	7.11 [4.56, 13.36]	5.27 [3.99, 8.33]	6.52 [3.00, 15.81]	4.87 [1.92, 8.56]
Prox. land. EVAR	—	—	—	4.92 [3.42, 6.24]	3.09 [1.39, 4.95]*	2.89 [2.13, 5.76]
Fabric EVAR	—	—	—	2.70 [1.10, 3.94]	2.89 [1.09, 5.26]	3.08 [1.19, 10.49]
Distal land. EVAR	—	—	—	1.83 [0.72, 3.10]	2.43 [0.99, 5.21]	2.48 [0.97, 6.30]
All landmarks	1.87 [0.90, 3.51]	2.21 [1.01, 4.08]	2.70 [1.39, 4.96]*	2.20 [0.82, 4.49]	2.46 [1.01, 4.77]	2.74 [1.16, 5.96]*

Table 4. Distance (mm) in axial view, represented as median [1st-3rd] quartiles, between landmarks annotated by two observers, and DL vs manual reference in both centers. Positive differences are a shift of DL predictions toward foot direction; negative differences are a shift of the DL predictions toward cranial direction.
* Significant difference (p<0.05) between interobserver difference and DL errors.
† Significant difference (p<0.05) between DL error in C1 and DL error in C2 data.
SMA: superior mesenteric artery; AAA: abdominal aortic aneurysm. C1: Training set (cross-validation). C2: External validation set.

Landmark	Pre-EVAR Interobserver variability	Pre-EVAR DL vs. manual C1	Pre-EVAR DL vs. manual C2	Post-EVAR Interobserver variability	Post-EVAR DL vs. manual C1	Post-EVAR DL vs. manual C2
Celiac trunk	-0.20 [-1.99, 0.79]	0.24 [-1.34, 1.56]†	1.61 [0.21, 3.18]*†	-0.06 [-1.60, 0.27]	0.10 [-1.04, 1.47]	0.78 [-0.70, 2.95]*
SMA	-0.71 [-2.35, 1.09]	0.21 [-1.45, 2.07]	1.61 [0.67, 2.83]*†	-0.15 [-1.70, 0.86]	0.06 [-1.37, 1.66]	1.30 [0.12, 2.39]*†
Renal arteries	-0.47 [-1.23, 0.34]	0.58 [-1.32, 1.83]*	1.55 [0.36, 2.54]*†	-0.53 [-1.30, 0.53]	0.56 [-1.08, 2.15]*	1.62 [0.56, 3.07]*†
Proximal origin AAA	-1.63 [-5.27, 1.76]	1.78 [-2.03, 5.02]*†	3.79 [2.18, 5.18]*†	-2.17 [-3.80, 0.36]	1.21 [-1.61, 4.30]*	1.50 [0.24, 4.48]*
Aortic bifurcation	1.91 [1.36, 2.91]	-2.47 [-3.90, -0.37]*†	-3.69 [-6.61, 1.11]*†	1.90 [0.07, 4.33]	-2.10 [-4.23, 0.00]*	-3.07 [-5.19, 0.15]*
Distal end AAA	-4.55 [-9.49, -1.65]	0.87 [-3.97, 6.47]*	2.42 [-3.93, 12.85]*	-5.27 [-8.33, -3.69]	3.05 [-2.42, 10.31]*	3.52 [-1.14, 7.73]*
Prox. land. EVAR	—	—	—	-4.92 [-6.24, -2.19]	0.96 [-1.80, 3.28]*	2.81 [1.44, 4.95]*†
Fabric EVAR	—	—	—	-2.49 [-3.75, -0.36]	1.06 [-1.16, 4.26]*	2.56 [0.07, 10.12]*
Distal land. EVAR	—	—	—	0.37 [-1.14, 2.00]	-0.14 [-2.27, 2.60]	-0.70 [-3.18, 1.65]
All landmarks	-0.30 [-2.72, 1.36]	0.09 [-2.15, 2.22]*	1.56 [-0.82, 3.68]*†	-0.91 [-3.72, 0.69]	0.35 [-2.08, 2.80]*	1.21 [-1.07, 3.81]*†

Environmental Impact

• Hardware Type:
  Training and validation of these models was done in a HPC node with two RTX 3090, 128 GB of VRAM

• Hours used:
  10 days for all models.

---

References

Garrido-Oliver, J., Tenezaca-Sari, X., Ribeiro, T. et al, Comprehensive Artificial Intelligence Based Analysis of Pre- and Postoperative Computed Tomography Scans in Patients Undergoing Endovascular Abdominal Aortic Repair. Eur J Vasc Endovasc Surg (2026). https://doi.org/10.1016/j.ejvs.2026.09.035.

Isensee, F., Jaeger, P.F., Kohl, S.A.A. et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods 18, 203–211 (2021). https://doi.org/10.1038/s41592-020-01008-z