Dynamic-Computed Tomography Angiography for Cerebral Vessel Templates and Segmentation

TL;DR

This study develops and evaluates arterial and venous angiographic templates derived from dynamic 4D-CTA to facilitate intracranial vessel segmentation on conventional CTA. It compares atlas-based, template-driven segmentation with deep learning methods trained on semi-automatic GTs generated via iCafe, using bone/subtracted and vessel-separated volumes to address arterial-venous separation. DL-based segmentation substantially outperforms atlas-based methods, especially for proximal arteries and major veins, across arterial, venous, and mixed phases, while template-based segmentation improves with voxel-labeled templates but struggles with distal vessels due to anatomical variability. The work demonstrates the feasibility and value of dynamic CTA-derived templates and DL approaches for reducing manual annotation and aiding radiologists in evaluating cerebrovascular conditions, with potential clinical impact in emergency and acute-care settings.

Abstract

Background: Computed Tomography Angiography (CTA) is crucial for cerebrovascular disease diagnosis. Dynamic CTA is a type of imaging that captures temporal information about the We aim to develop and evaluate two segmentation techniques to segment vessels directly on CTA images: (1) creating and registering population-averaged vessel atlases and (2) using deep learning (DL). Methods: We retrieved 4D-CT of the head from our institutional research database, with bone and soft tissue subtracted from post-contrast images. An Advanced Normalization Tools pipeline was used to create angiographic atlases from 25 patients. Then, atlas-driven ROIs were identified by a CT attenuation threshold to generate segmentation of the arteries and veins using non-linear registration. To create DL vessel segmentations, arterial and venous structures were segmented using the MRA vessel segmentation tool, iCafe, in 29 patients. These were then used to train a DL model, with bone-in CT images as input. Multiple phase images in the 4D-CT were used to increase the training and validation dataset. Both segmentation approaches were evaluated on a test 4D-CT dataset of 11 patients which were also processed by iCafe and validated by a neuroradiologist. Specifically, branch-wise segmentation accuracy was quantified with 20 labels for arteries and one for veins. DL outperformed the atlas-based segmentation models for arteries (average modified dice coefficient (amDC) 0.856 vs. 0.324) and veins (amDC 0.743 vs. 0.495) overall. For ICAs, vertebral and basilar arteries, DL and atlas -based segmentation had an amDC of 0.913 and 0.402, respectively. The amDC for MCA-M1, PCA-P1, and ACA-A1 segments were 0.932 and 0.474, respectively. Conclusion: Angiographic CT templates are developed for the first time in literature. Using 4D-CTA enables the use of tools like iCafe, lessening the burden of manual annotation.

Figures (8)

• Figure 1: Panel (a) Dynamic CTA from a sample patient. The axis on the left shows the acquisition time of each phase; the red and blue arrows point to the arterial phase and venous phase images respectively. Panel (b) provides an overview of the cohort selection process for constructing the template and training the deep learning model.
• Figure 2: Flowchart illustrating the evaluation data selection.From an initial pool of 1937 scans performed on the institution’s 4D-CT scanner, we identified 140 patients with Dynamic CTA (having at least two phases). After excluding those lacking subtracted images, 68 patients remained; of these, 11 provided subtracted CTA suitable for semi-automatic artery/vein annotation.
• Figure 3: (a) Preprocessing steps for a dynamic CTA from a single patient. For both phases, a head ROI is used on the CTA and the subtracted CTA to suppress signals from outside the skull. Next, the subtracted images are processed using Algorithm 1 to create the vessel-separated images. (b) Shows the workflow used to identify the optimal template. From all patients from the template cohort and for each group of images $X_a,X_v,S_a^*,$ and $S_v^*$, the pipeline creates intermediate templates $T_{a,k},T_{v,k},T_{as,k},$ and $T_{vs,k}$ respectively, at iteration $k$. After convergence, the identified optimal template is labelled as $T_{a},T_{v},T_{as}^*,$ and $T_{vs}^*$, respectively. (c) Shows how the evaluation of the registration-based segmentation is conducted with the evaluation cohort. For both arterial and venous ROIs, we test three groups of registrations with different combinations of moving images and fixed images. The identified non-linear transformation is then applied to the corresponding ROI.
• Figure 4: Dynamic CTA for all patients are collected (Step (a)). The subtracted images are processed to suppress voxels outside the head and resample to 0.468mm$^3$ isotropic resolution. Finally, the vessel-separated, artery-only ($S_a^*$) and vein-only ($S_v^*$) are computed (Step (b)). The vessel-separated images are then processed individually using iCafe's vessel tracing algorithm. These traces are then validated and exported for model training (Step (c)). In Step (d), for the model training, all available CTA data (i.e., 3-19 volumes per patient) is paired with the corresponding ground truth segmentation, and a deep learning model is trained.
• Figure 5: An overview of extracting the ground truth segmentations of veins and arteries using iCafe. Traces are first identified for both vessels. For the artery-only volume, the traces are edited for accuracy and exported into segmentations. For the veins, the vessel traces are directly exported into segmentations. Distal branches from the venous segmentation were removed using 3D Slicer.