Anatomy- And Time-Based Post-Processing with CT Perfusion-Derived Time–Vessel Maps to Reduce Incorrect Occlusion Detections in Late-Phase CT Angiography

Copyright German Cancer Research Center (DKFZ) and contributors.

Background/Objectives: Computer-aided vessel occlusion detection for acute ischemic stroke is typically developed on early-phase CT angiography (CTA). In late-phase CTA, enhanced venous contrast can mimic arterial occlusions and lead to incorrect occlusion detections. To counteract this, we propose a constraint-based post-processing framework that leverages anatomical and temporal vascular information from CT perfusion (CTP) scans. Methods: Vascular anatomy and relative contrast-arrival times were encoded as time–vessel maps. Three complementary constraints worked on these maps to remove incorrect detections based on distance-to-vessel truncation points, contrast-arrival time, and spatial priors. Constraints were fixed on a 32-case late-phase development subset and applied to an external held-out cohort of late-phase CTA (N = 354). To enable application without CTP, an auxiliary CTA-to-time–vessel map generator was designed to estimate maps directly from CTA. Results: In the external cohort, the framework reduced incorrect occlusions per scan from 3.9 (95% CI 3.6–4.2) to 2.1 (95% CI 1.9–2.3), preserving occlusion-level sensitivity at 80% (95% CI 69–91). Patient-level AUROC was 83 for the baseline versus 88 with post-processing. Improvements were also observed in detectors exposed to late-phase CTA, where post-processing reduced incorrect occlusions per scan from 2.8 to 1.9. As only 51 occlusion-positive cases were present in the external cohort, confidence intervals were wide. Consequently, these estimates should be interpreted with caution. Conclusions: Anatomically and temporally informed post-processing reduced incorrect occlusion detections in late-phase CTA without measurable loss of sensitivity, potentially improving robustness against CTA phase changes in external institutions.

Please cite the following paper if you use this code: (https://www.mdpi.com/2075-4418/16/18/2972)

Martínez Mora, A.; Mojtahedi, M.; de Vries, L.; Baumgartner, M.; Kirchhoff, Y.; Zenk, M.; Eckstein, K.; Kächele, J.; Brugnara, G.; Bendszus, M.; et al. Anatomy- and Time-Based Post-Processing with CT Perfusion-Derived Time–Vessel Maps to Reduce Incorrect Occlusion Detections in Late-Phase CT Angiography. Diagnostics 2026, 16, 2972. https://doi.org/10.3390/diagnostics16182972

Installation

All of the code in this repository runs from two conda environments:

• ctp-postprocess — the main environment: PyTorch, nnDetection (occlusion detection), the Skeleton-recall nnU-Net v2 fork bundled in this repo (CTA-to-time–vessel map generator), SimpleElastix (registration), and other dependencies.
• ctp-totalseg — an environment containing only TotalSegmentator and its own pinned nnunetv2. It is kept separate because TotalSegmentator's PyPI nnunetv2 would otherwise collide with the custom Skeleton-recall fork used everywhere else.

Note: nnDetection's official releases are only tested against PyTorch 1.x (its README states PyTorch 2.0+ is not supported), while Skeleton-recall requires torch>=2.1.2. The steps below install nnDetection against PyTorch 2.x anyway, overriding its pinned dependencies. This combination has been built and smoke-tested end to end (CUDA extension compiles and runs, box_iou_np/load_pickle import correctly) on Python 3.10 + torch 2.5.1+cu118 + an RTX 4090 — but nnDetection's own Lightning-based training/inference pipeline (nndet_train, nndet_predict) was not exercised, only the lightweight nndet.io / nndet.core.boxes utilities this repo's scripts actually import.

1. Main environment (ctp-postprocess)

conda create -n ctp-postprocess python=3.10
conda activate ctp-postprocess

# --- PyTorch (CUDA 11.8 build; adjust to your driver/CUDA toolkit) ---
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 \
    --index-url https://download.pytorch.org/whl/cu118

# --- Build tools needed to compile nnDetection's CUDA extensions ---
conda install -y -c conda-forge gxx_linux-64 ninja
conda install -y -c nvidia/label/cuda-11.8.0 cuda-toolkit

# --- Direct dependencies of nnDetection + this repo's top-level scripts ---
pip install -r requirements.txt
pip install hydra-core --upgrade --pre

# --- nnDetection. In the future to be updated to nnDetection v2, when this repository is out ---
git clone https://github.com/MIC-DKFZ/nnDetection.git
cd nnDetection

# Install nnDetection with argument --no-build-isolation:
FORCE_CUDA=1 CUDA_HOME=$CONDA_PREFIX TORCH_CUDA_ARCH_LIST="<your GPU's compute capability, e.g. 8.9 for RTX 4090>" \
    pip install -v -e . --no-build-isolation --no-deps
cd ..

# --- Skeleton-recall (custom nnU-Net v2 fork bundled in this repo for CTA-to-time-vessel map generator) ---
pip install -e ./Skeleton-recall

# --- SimpleElastix build of SimpleITK, for registration capabilities
pip install SimpleITK-SimpleElastix

nnDetection also requires a few environment variables to be set:

export det_data=/path/to/det_data # Folder with CTA data and vessel occlusion labels
export det_models=/path/to/det_models # Folder where to store vessel occlusion detectors
export OMP_NUM_THREADS=1
export det_num_threads=6

2. TotalSegmentator environment (ctp-totalseg)

conda create -n ctp-totalseg python=3.10
conda activate ctp-totalseg

# Pin torch to a CUDA build your driver actually supports BEFORE installing
# TotalSegmentator
# Check your driver's max supported CUDA version with `nvidia-smi` and adjust the
# --index-url below (cu118/cu121/cu124/...) accordingly.
pip install torch==2.5.1 torchvision==0.20.1 --index-url https://download.pytorch.org/whl/cu124

pip install -r requirements-totalseg.txt

postprocess_end2end.py calls TotalSegmentator as an external subprocess and does not need ctp-totalseg to be active — just point it at the environment's binary:

export TOTALSEG_BIN=/path/to/conda/envs/ctp-totalseg/bin/TotalSegmentator
# or pass --totalseg_bin /path/to/conda/envs/ctp-totalseg/bin/TotalSegmentator

Main commands

If you already have a trained vessel occlusion detector in nnDetection and a trained CTA-to-time-vessel map generator

To access our own nnDetection and generator trained models, contact us for sharing at reasonable enquiry

conda activate ctp-postprocess
python postprocess_end2end.py --d /path/to/cta/data --m /path/to/generator/model --p /path/to/det_models/TaskXYZ/model_name/foldF/val_or_test_predictions --o /path/to/postprocessed/predictions --cfg fpr_cfg.json --brain_cache /path/to/store/brain/segmentations --totalseg_bin /path/to/conda/envs/ctp-totalseg/bin/TotalSegmentator 

If you want to start from a folder with CTP data and a folder with CTA data. Required data structure:

ctp_folder/
├── case0/
│   ├── case0_t_00.nii.gz
│   ├── case0_t_01.nii.gz
│   ├── ...
│   └── case0_t_NN.nii.gz
├── ...
└── caseM/
    ├── caseM_t_00.nii.gz
    ├── caseM_t_01.nii.gz
    ├── ...
    └── caseM_t_NN.nii.gz

ctp_folder_with_time_information/
├── case0_AcquisitionDateTime.npy
├── ...
└── caseM_AcquisitionDateTime.npy

cta_folder/
├── case0.nii.gz
├── ...
└── caseM.nii.gz

1. CTP-to-CTA registration

conda activate ctp-postprocess
python register.py --cta /your/cta/folder --ctp /your/ctp/folder --out /registered/ctp/folder

2. CTP curation (start with the output registered CTP folder from Step 1).

USE THE TOTALSEG ENV!!

It requires a nnU-Net vessel segmentation model, with a folder path in 'config.json' ('mca_cpt'). Right now this field is named as PLACEHOLDER

Contact us if you require the model.

conda activate ctp-totalseg
python preprocessing.py --folder /registered/ctp/folder --time /ctp/time/folder --out /curated/ctp/folder

If CTP time files (.npy) require resampling after CTP image information curation, run:

conda activate ctp-postprocess
python utils/resample_time_info.py --time /ctp/time/folder --cta /your/cta/folder --out /resampled/ctp/time/folder

3. Time-vessel map extraction (start with the curated CTP folder from Step 2).

USE THE TOTALSEG ENV!! Brain segmentations are outputted in the folder under 'skull' argument If you have resampled your CTP time files, use the path /resampled/ctp/time/folder for the --time argument

conda activate ctp-totalseg
python extract_tta.py --ctp /curated/ctp/folder --cta /your/cta/folder --time /ctp/time/folder --skull /folder/where/to/store/brain/segmentations --tta /folder/with/time-vessel-maps

4. Postprocessing constraints application

If you have to use the CTA-to-time-vessel map generator model

conda activate ctp-postprocess
python postprocess_end2end.py --d /path/to/cta/data --m /path/to/generator/model --p /path/to/det_models/TaskXYZ/model_name/foldF/val_or_test_predictions --o /path/to/postprocessed/predictions --cfg fpr_cfg.json --brain_cache /path/to/store/brain/segmentations --totalseg_bin /path/to/conda/envs/ctp-totalseg/bin/TotalSegmentator 

If you have already-derived time-vessel maps

conda activate ctp-postprocess
python postprocess_end2end.py --d /path/to/cta/data --m /path/to/generator/model --t /folder/with/time-vessel-maps --p /path/to/det_models/TaskXYZ/model_name/foldF/val_or_test_predictions --o /path/to/postprocessed/predictions --cfg fpr_cfg.json --brain_cache /path/to/store/brain/segmentations --totalseg_bin /path/to/conda/envs/ctp-totalseg/bin/TotalSegmentator 

5. Development of CTA-to-time-vessel map generator

5.1 Distance map computation

conda activate ctp-postprocess
python compute_distance.py --t /folder/with/time-vessel-maps --b /folder/where/to/store/brain/segmentations --o /folder/with/distance-maps

5.2 Data preparation

conda activate ctp-postprocess
python utils/prepare_raw_data_generator.py --c /your/cta/folder --t /folder/with/time-vessel-maps --d /folder/with/distance-maps --b /folder/where/to/store/brain/segmentations --task task_name (nnUNet style, "DatasetXYZ") --o /folder/with/raw/generator/data

Adapted nnU-Net setup

nnU-Net also requires a few environment variables to be set:

export nnUNet_raw=/folder/with/raw/generator/data/raw_cropped # Folder with processed raw data by utils/prepare_raw_generator.py
export nnUNet_preprocessed=/folder/with/nnUNet/preprocessed/data # Folder where to store vessel occlusion detectors
export nnUNet_results=/folder/with/generator/model

5.3 Preprocessing

conda activate ctp-postprocess
nnUNetv2_extract_fingerprint -d TASK_ID
nnUNetv2_plan_experiment -d TASK_ID -pl nnUNetPlannerResEncL -preprocessor_name MultiChannelSegPreprocessor
nnUNetv2_preprocess -d TASK_ID -plans_name nnUNetResEncUNetLPlans -c 3d_fullres

5.4 Training (assuming some separate test set exists)

conda activate ctp-postprocess
nnUNetv2_train TASK_ID 3d_fullres all -p nnUNetResEncUNetLPlans -tr nnUNetRegressionTrainer --use_compressed

5.5 Inference

If test data is raw and has not been processed by prepare_raw_data_generator.py, prepare inference data by masking out non-brain voxels with -1024 HU

conda activate ctp-postprocess
python utils/obtain_brainProd_images.py --i /raw/CTA/folder --b /folder/where/to/store/brain/segmentations --o /folder/with/processed/inference/data

nnU-Net inference

conda activate ctp-postprocess
python Skeleton-Recall/nnunetv2/utilities/predict_folder.py --d /folder/with/processed/inference/data --o /folder/with/output/predictions --m /folder/with/generator/model/DatasetXYZ/nnUNetRegressionTrainer__nnUNetResEncUNetLPlans__3d_fullres --b /folder/where/to/store/brain/segmentations --cfg params_generator.json

5.6 Inference evaluation

conda activate ctp-postprocess
python Skeleton-recall/nnunetv2/evaluation/evaluate_predictions.py --folder_ref /folder/with/ground-truth/time-vessel-maps --folder_pred /folder/with/output/predictions --output_file /output/metric/file.json