A vascular synthetic model for improved aneurysm segmentation and detection via Deep Neural Networks

Abstract: We hereby present a full synthetic model, able to mimic the various constituents of the cerebral vascular tree: the cerebral arteries, the bifurcations and the intracranial aneurysms. By building this model, our goal was to provide a substantial dataset of brain arteries which could be used by a 3D Convolutional Neural Network (CNN) to either segment or detect/recognize various vascular diseases (such as artery dissection/thrombosis) or even some portions of the cerebral vasculature, such as the bifurcations or aneurysms. In this study, we will particularly focus on Intra-Cranial Aneurysm (ICA) detection and segmentation. The cerebral aneurysms most often occur on a particular structure of the vascular tree named the Circle of Willis. Various studies have been conducted to detect and monitor the ICAs and those based on Deep Learning (DL) achieve the best performances. Specifically, in this work, we propose a full synthetic 3D model able to mimic the brain vasculature as acquired by Magnetic Resonance Angiography (MRA), and more particularly the Time Of Flight (TOF) principle. Among the various MRI modalities, the MRA-TOF allows to have a relatively good rendering of the blood vessels and is non-invasive (no contrast liquid injection). Our model has been designed to simultaneously mimic the arteries geometry, the ICA shape and the background noise. The geometry of the vascular tree is modeled thanks to an interpolation with 3D Spline functions, and the statistical properties of the background MRI noise is collected from MRA acquisitions and reproduced within the model. In this work, we thoroughly describe the synthetic vasculature model, we build up a neural network designed for ICA segmentation and detection, and finally, we carry out an in-depth evaluation of the performance gap gained thanks to the synthetic model data augmentation.

• The paper’s main contribution is a synthetic model that generates realistic 3D cerebral vasculature and aneurysms to improve CNN-based intracranial aneurysm detection.
• It integrates precise 3D spline-based arterial geometry and MRA-TOF noise simulation to mimic the structural variability of human brain vessels.
• Experimental results show that augmenting real data with synthetic patches increases lesion sensitivity from 75.6% to 88.97%, with a modest rise in false positives.

Synthetic Vascular Model for Aneurysm Detection

This paper introduces a comprehensive synthetic model designed to mimic the cerebral vascular tree, including arteries, bifurcations, and intracranial aneurysms (ICAs). The primary goal is to generate a substantial dataset of brain arteries for training 3D Convolutional Neural Networks (CNNs) to segment and detect vascular diseases, specifically ICAs, using Magnetic Resonance Angiography (MRA) Time-Of-Flight (TOF) imaging. The model aims to address the challenge of limited annotated datasets in medical imaging by providing a fully synthetic, customizable data source.

Model Design and Implementation

The synthetic vasculature model (Figure 1) consists of three main components: the geometry of the arteries, the surrounding MRA-TOF noise, and the modeled aneurysm.

Figure 1: Overview of the global procedure, encompassing the training step using the synthetic images and the inference step.

Arteries Geometry

The geometry of the vascular tree is modeled using 3D spline functions, allowing for precise control over the shape, orientation, diameter, and tortuosity of the arteries. The process involves extracting 3D coordinates of the arteries' skeleton from segmented MRA-TOF volumes and fitting these centerlines using 3D splines. The polynomial coefficients of the splines are then altered to introduce slight distortions, mimicking the structural variability observed in human vasculature. The diameters of the arteries are determined using a vascular tree characterization tool, and the arteries are thickened by convolving the centerlines with a spherical kernel, the size of which is adapted to the corresponding diameter. Geometric distortions (elastic deformations) are applied to the 3D kernel before convolution to create realistic artery shapes. The model also allows for setting a target gray level amplitude for the arteries and introducing inhomogeneities by applying various kernels along the vessel's centerline.

MRA-TOF Noise

The model includes a detailed representation of the brain, composed of fluids and white/gray matter, all affected by background noise. The different brain components are separated using multi-threshold segmentation or Gaussian Mixture Models (GMMs). Each component can then be geometrically distorted before replicating its overlaying noise. The noise generation process involves creating a high-frequency Gaussian noise with an average set to the target 3D crop, which is then smoothed using a Gaussian filter to match the statistical properties of the target MRA-TOF portion.

Aneurysm Modeling

Synthetic aneurysms are incorporated into the model by first creating a simple 3D sphere, which is then distorted using elastic deformations. The aneurysm center is aligned onto the bisector between the two daughter arteries, with the distance from the aneurysm center to the bifurcation node computed based on the aneurysm radius, the average radius of the branches forming the bifurcation, and the angle formed by the two daughter arteries. A growth parameter allows for adjusting the shift from the aneurysm center to the vessel wall, enabling the modeling of various growth stages.

Figure 2 and Figure 3 show examples of the model's output, demonstrating the accurate representation of both bifurcations and aneurysms.

Figure 2: Comparison between the modeled bifurcations and the Ground Truth crop from a MRA-TOF. We show the comparison on terms of both gray level voxels patches (leftmost panels) and 3D bifurcation layout (rightmost panels).

Figure 3: Comparison between the modeled bifurcations (bearing an aneurysm). In the upper sub-figures (3D representations), the aneurysm is represented in blue, the mother artery in green.

Experimental Evaluation

The synthetic model's effectiveness was evaluated by training a 3D U-Net CNN for ICA segmentation and detection. The dataset consisted of 190 MRA-TOFs scans of unruptured ICAs, divided into training and test sets. The training set was augmented with 998 synthetic patches generated by the model. Two experiments were conducted: one using only real MRA-TOF patches (Exp.#1) and another using a combination of real and synthetic patches (Exp.#2).

The results showed that the CNN trained with both real and synthetic patches (Exp.#2) achieved a significantly improved lesion-level sensitivity of $88.97\%$ compared to the CNN trained solely on real data (Exp.#1), which had a sensitivity of $75.60\%$ (Figure 4).

Figure 4: Detection performance of the CNN using real data vs real and synthetic data.

The patient-level sensitivity also improved from $79.65\%$ in Exp.#1 to $91.36\%$ in Exp.#2. However, the false positive rate increased slightly from $0.22$ in Exp.#1 to $0.40$ in Exp.#2. The model's performance varied depending on the size and location of the aneurysms. Exp.#2 showed improved detection rates for small aneurysms (less than 2 mm) and for aneurysms located along the Middle Cerebral Artery (MCA) and Internal Carotid Artery.

Figure 5 shows the number of missed detections with respect to the aneurysms positions in the test dataset.

Figure 5: Missed detections with respect to the aneurysms positions in the test dataset.

The Dice score for the detected aneurysms was comparable between the two experiments, with an average of 0.7585 for Exp.#1 and 0.7613 for Exp.#2 (Figure 6).

Figure 6: Dice similarity coefficient of true ICAs.

Conclusion

The synthetic vasculature model provides a valuable tool for generating large datasets of realistic brain arteries and aneurysms, which can be used to train CNNs for ICA segmentation and detection. The results demonstrate that augmenting real MRA-TOF data with synthetic data can significantly improve the sensitivity of ICA detection, particularly for small aneurysms and aneurysms located in specific regions of the brain. While the addition of synthetic data may lead to a slight increase in the false positive rate, the overall improvement in sensitivity makes this approach a promising tool for assisting neuroradiologists in the detection of intracranial aneurysms. Future work may focus on refining the synthetic model to further reduce the false positive rate and to extend its applicability to other imaging modalities, such as CTA and DSA. The source code for the synthetic Vascular Models (VaMos) has been made available on a GitLab repository.