Anatomical labeling of intracranial arteries with deep learning in patients with cerebrovascular disease

Generative modeling of the Circle of Willis using 3D-StyleGAN

The circle of Willis (CoW) is a network of cerebral arteries with significant inter-individual anatomical variations. Deep learning has been used to characterize and quantify the status of the CoW in various applications for the diagnosis and treatment of cerebrovascular disease. In medical imaging, the performance of deep learning models is limited by the diversity and size of training datasets. To address medical data scarcity, generative AI models have been applied to generate synthetic vessel neuroimaging data. However, the proposed methods produce synthetic data with limited anatomical fidelity or downstream utility in tasks concerning vessel characteristics. We adapted the StyleGANv2 architecture to 3D to synthesize Time-of-Flight Magnetic Resonance Angiography (TOF MRA) volumes of the CoW. For generative modeling, we used 1782 individual TOF MRA scans from 6 open source datasets. To train the adapted 3D StyleGAN model with limited data we employed differentiable data augmentations, used mixed precision and a cropped region of interest of size 32 × 128 × 128 to tackle computational constraints. The performance was evaluated quantitatively using the Fréchet Inception Distance (FID), MedicalNet distance (MD) and Area Under the Curve of the Precision and Recall Curve for Distributions (AUC-PRD). Qualitative analysis was performed via a visual Turing test. We demonstrated the utility of generated data in a downstream task of multiclass semantic segmentation of CoW arteries. Vessel segmentation performance was assessed quantitatively using the Dice coefficient and the Hausdorff distance. The best-performing 3D StyleGANv2 architecture generated high-quality and diverse synthetic TOF MRA volumes (FID: 12.17, MD: 0.00078, AUC-PRD: 0.9610). Multiclass vessel segmentation models trained on synthetic data alone achieved comparable performance to models trained using real data in most arteries. The addition of synthetic data to a baseline training set improved segmentation performance in underrepresented artery segments, similar to the addition of real data. In conclusion, generative modeling of the Circle of Willis via synthesis of 3D TOF MRA data paves the way for generalizable deep learning applications in cerebrovascular disease. In the future, the extensions of the provided methodology to other medical imaging problems or modalities with the inclusion of pathological datasets has the potential to advance the development of more robust AI models for clinical applications.

A novel procedure for medial axis reconstruction of vessels from Medical Imaging segmentation

A procedure for reconstructing the central axis from diagnostic image processing is presented here, capable of solving the widespread problem of stepped shape effect that characterizes the most common algorithmic tools for processing the central axis for diagnostic imaging applications through the development of an algorithm correcting the spatial coordinates of each point belonging to the axis from the use of a common discrete image skeleton algorithm. The procedure is applied to the central axis traversing the vascular branch of the cerebral system, appropriately reconstructed from the processing of diagnostic images, using investigations of the local intensity values identified in adjacent voxels. The percentage intensity of the degree of adherence to a specific anatomical tissue acts as an attraction pole in the identification of the spatial center on which to place each point of the skeleton crossing the investigated anatomical structure. The results were shown in terms of the number of vessels identified overall compared to the original reference model. The procedure demonstrates high accuracy margin in the correction of the local coordinates of the central points that permits to allocate precise dimensional measurement of the anatomy under examination. The reconstruction of a central axis effectively centered in the region under examination represents a fundamental starting point in deducing, with a high margin of accuracy, key informations of a geometric and dimensional nature that favours the recognition of phenomena of shape alterations ascribable to the presence of clinical pathologies.

Automated anatomical labeling of the intracranial arteries via deep learning in computed tomography angiography

Background and purpose: Anatomical labeling of the cerebral vasculature is a crucial topic in determining the morphological nature and characterizing the vital variations of vessels, yet precise labeling of the intracranial arteries is time-consuming and challenging, given anatomical structural variability and surging imaging data. We present a U-Net-based deep learning (DL) model to automatically label detailed anatomical segments in computed tomography angiography (CTA) for the first time. The trained DL algorithm was further tested on a clinically relevant set for the localization of intracranial aneurysms (IAs). Methods: 457 examinations with varying degrees of arterial stenosis were used to train, validate, and test the model, aiming to automatically label 42 segments of the intracranial arteries [e.g., 7 segments of the internal carotid artery (ICA)]. Evaluation metrics included Dice similarity coefficient (DSC), mean surface distance (MSD), and Hausdorff distance (HD). Additionally, 96 examinations containing at least one IA were enrolled to assess the model's potential in enhancing clinicians' precision in IA localization. A total of 5 clinicians with different experience levels participated as readers in the clinical experiment and identified the precise location of IA without and with algorithm assistance, where there was a washout period of 14 days between two interpretations. The diagnostic accuracy, time, and mean interrater agreement (Fleiss' Kappa) were calculated to assess the differences in clinical performance of clinicians. Results: The proposed model exhibited notable labeling performance on 42 segments that included 7 anatomical segments of ICA, with the mean DSC of 0.88, MSD of 0.82 mm and HD of 6.59 mm. Furthermore, the model demonstrated superior labeling performance in healthy subjects compared to patients with stenosis (DSC: 0.91 vs. 0.89, p < 0.05; HD: 4.75 vs. 6.19, p < 0.05). Concurrently, clinicians with model predictions achieved significant improvements when interpreting the precise location of IA. The clinicians' mean accuracy increased by 0.04 (p = 0.003), mean time to diagnosis reduced by 9.76 s (p < 0.001), and mean interrater agreement (Fleiss' Kappa) increased by 0.07 (p = 0.029). Conclusion: Our model stands proficient for labeling intracranial arteries using the largest CTA dataset. Crucially, it demonstrates clinical utility, helping prioritize the patients with high risks and ease clinical workload.

Graph convolutional networks for automated intracranial artery labeling

Purpose

Unruptured intracranial aneurysms (UIAs) can cause aneurysmal subarachnoid hemorrhage, a severe and often lethal type of stroke. Automated labeling of intracranial arteries can facilitate the identification of risk factors associated with UIAs. This study aims to improve intracranial artery labeling using atlas-based features in graph convolutional networks.

Approach

We included three-dimensional time-of-flight magnetic resonance angiography scans from 150 individuals. Two widely used graph convolutional operators, GCNConv and GraphConv, were employed in models trained to classify 12 bifurcations of interest. Cross-validation was applied to explore the effectiveness of atlas-based features in node classification. The results were tested for statistically significant differences using a Wilcoxon signed-rank test. Model repeatability and calibration were assessed on the test set for both operators. In addition, we evaluated model interpretability and node feature contribution using explainable artificial intelligence.

Results

Atlas-based features led to statistically significant improvements in node classification (p < 0.05 ). The results showed that the best discrimination and calibration performances were obtained using the GraphConv operator, which yielded a mean recall of 0.87, precision of 0.90, and expected calibration error of 0.02.

Conclusions

The addition of atlas-based features improved node classification results. The GraphConv operator, which incorporates higher-order structural information during training, is recommended over the GCNConv operator based on the accuracy and calibration of predicted outcomes.