A Rupture Risk Prediction Model for Cerebral Aneurysms Combining Hemodynamic and Morphological Parameters

Abstract

Cerebral aneurysms are a serious health concern, and their treatment prior to rupture is crucial to avoid severe disability. Thus, the accurate prediction of rupture risk is of paramount importance, and various hemodynamic factors have been utilized for this purpose. In this dissertation, novel methods of predicting rupture risk in cerebral aneurysms are proposed through artificial intelligences trained using both hemodynamic and morphological parameters, which will aid in better decision-making about treatment. For accomplish this, the computational analyses, computational fluid dynamics (CFD) and fluid-structure interaction (FSI), used to calculate the hemodynamic parameters of cerebral aneurysms. In addition, new morphological parameters that can evaluate irregularities of cerebral aneurysms were devised. Before developing the rupture risk prediction models, standard criteria for computational analyses were established because the analysis conditions for computational analyses are chosen without any specific criteria, which results in inconsistencies of computational analyses depending on the practitioner. In CFD, the changes of hemodynamic parameters depending on the inlet boundary conditions, plug and Womersley flow, and the outlet boundary conditions, zero and pulsatile pressures, were investigated. In addition, the difference in the assumption of blood viscosity was also analyzed with respect to the flow rate. It was confirmed that if the entrance length was sufficiently secured, the inlet and outlet boundary conditions did not affect the CFD results. Furthermore, it was observed that the difference in the hemodynamic parameters between Newtonian and non-Newtonian fluid decreased as the flow rate increased. In addition, the effect of mechanical property and wall thickness variations of cerebral arteries and boundary conditions of FSI on the hemodynamic parameters were evaluated. There was no significant difference between linear and non-linear material models of mechanical properties. Also, although there was numerical error in element types such as shell and solid element for discretizing the computational model of a cerebral artery, it could be terminated by normalizing the results of hemodynamic parameters. Moreover, there were no significant differences between the results depending on the constraint conditions at ends of the blood vessel model. Additionally, the experimental validations of CFD and FSI were performed. The artificial circle of Willis model was fabricated using liquid-assisted dip coating method. The flow circulation system with the artificial blood vessel were then developed, and fluid flows were visualized by injecting the colored inks. Furthermore, the inflation test with the artificial blood vessel of idealized aneurysm model was performed to validate the FSI results by evaluating the amount of deformation of the blood vessel depending on the internal pressure. Finally, it was confirmed that the CFD and FSI results was good agreement with the experimental results. After establishing the standard criteria for conducting the computational analyses, the relationship between aneurysm formation and hemodynamic parameters was investigated to evaluate whether the hemodynamic parameters are useful parameters related to the mechanism of intracranial aneurysms. A total of thirty-seven intracranial paraclinoid aneurysms were enrolled for quantitative comparison. The locations of high wall shear stress (WSS) and high strain were extracted from the CFD and FSI, respectively. The distances between the aneurysm formation site and the locations of high WSS or high strain were then calculated. The average distance from the location of the aneurysm formation site to the high strain was smaller than the average distance to the high WSS. Not only high WSS but also high strain was an important hemodynamic parameter related to the formation of cerebral aneurysms. In addition, fifty-eight patients with paraclinoid aneurysms on one side and two patients with paraclinoid aneurysms on both sides were enrolled to investigate the reason why aneurysms do not initiate in non-aneurysmal arteries. Using magnetic resonance angiography, the left and right internal carotid arteries (ICAs) of each patient were reconstructed. For patients with an aneurysm on one side, the ICA with the paraclinoid aneurysm was defined as the aneurysmal artery after eliminating the aneurysm, while the opposite ICA without an aneurysm was defined as the non-aneurysmal artery. CFD and FSI were conducted for both aneurysmal and non-aneurysmal arteries to determine the relationship between high hemodynamic parameters and aneurysm formation site. The results showed that high WSS and strain locations were well-matched with the aneurysm formation site for aneurysmal arteries, and there were considerable correlations between high WSS and strain locations. However, there was no significant relationship between high WSS and strain locations for non-aneurysmal arteries. For the two cases with paraclinoid aneurysms on both ICAs, every aneurysm formed in the vicinity of the location where both high WSS and strain were high. These investigations showed that the hemodynamic parameters were highly related to the mechanism of cerebral aneurysms. After confirming the correlation between hemodynamic parameters and cerebral aneurysm, rupture risk prediction models were constructed using deep learning based on the convolutional neural network (CNN) that can evaluate the complex relationship of hemodynamic parameters between ruptured and unruptured cerebral aneurysm, and total of 123 cerebral aneurysms were enrolled. Hemodynamic parameters, WSS and strain, were first calculated using CFD and FSI, and then converted into images for training the CNN using a novel approach. In particular, new data augmentation methods were devised to obtain sufficient training data. A total of 53,136 images generated by data augmentation were used to train and test the CNN. The CNNs trained with WSS, strain, and combination images (which have both information of WSS and strain) had area under the receiver operating characteristics (AUROC) curve values of 0.716, 0.741, and 0.883, respectively. Based on the cut-off values, the CNN trained with WSS or strain images alone was not highly predictive. However, the CNN trained with combination images of WSS and strain showed a comparable sensitivity and specificity of 0.81 and 0.82, respectively. In addition, new morphological parameters were devised based on the mass moment of inertia to improve the practicality of the rupture risk prediction models in clinical applications. Morphological parameters have been widely employed to assess the rupture risk of intracranial aneurysms, owing to their high predictive capacity. However, the current morphological parameters are limited in their ability to evaluate the irregularities of intracranial aneurysms. The mass moment of inertia made it possible to quantitatively evaluate the irregularities of cerebral aneurysms that could not considered with the conventional morphological parameters. The mass moment of inertia was normalized using the aneurysm mass, aneurysm neck area, and parent vessel area, resulting in a dimensionless parameter to evaluate relative differences of morphologies in various cerebral aneurysms. The hemodynamic parameters and normalized mass moment of inertia were calculated for 125 intracranial aneurysms (80 unruptured and 45 ruptured aneurysms), and artificial neural networks (ANN) trained with each parameter were utilized for rupture risk prediction. When the rupture risk prediction model was trained using both normalized mass moment of inertia and hemodynamic parameters, the highest performance for rupture risk prediction was obtained (sensitivity, 85.2 %; specificity, 92.9 %; AUROC of curve, 0.941). Despite several limitations, this dissertation suggested that the rupture risk prediction models combining the hemodynamic and morphological parameters is feasible. Therefore, the hemodynamic parameters and mass moment of inertia proposed in this dissertation may serve as significant factors for assessing the risk of aneurysm rupture, in conjunction with the clinical features. However, additional studies are needed to facilitate the application of the approaches devised in this dissertation and to promote their widespread use in clinical practice. |뇌동맥류는 파열 시 심각한 장애를 일으키므로, 뇌동맥류가 파열되기 전에 이를 치료해야 한다. 따라서, 뇌동맥류의 파열 위험성을 정확하게 예측하는 것이 무엇보다 중요하며, 이를 위해 다양한 혈류역학적 인자 (Hemodynamic parameters)들이 활용되고 있다. 본 논문에서는 혈류역학적 인자와 형상학적 인자 (Morphological parameter)를 기반으로 학습된 인공지능 모델들을 통해 뇌동맥류 파열 위험성을 예측하는 새로운 방법을 제안하였다. 뇌동맥류의 혈류역학적 인자를 계산하기 위해 전산 유체 역학 (Computational fluid dynamics) 및 유체-구조 연성 (Fluid-structure interaction) 해석들을 활용하였다. 또한 뇌동맥류 형상의 불규칙도 (Irregularity)을 정량적으로 평가할 수 있는 새로운 형상학적 인자를 고안하였다. 먼저, 뇌동맥류 파열 위험성 예측 모델을 개발하기 전에 뇌동맥류의 혈류역학적 인자들을 평가하기 위한 전산 혈류 해석의 기준을 수립하였다. 경계 조건 및 유체 특성 변화에 대한 혈류 역학적 인자들의 차이를 분석하였고, 유체-구조 연성 해석에 사용되는 경계 조건들 및 뇌동맥의 Young’s modulus 및 벽 두께 변화가 혈류역학적 인자에 미치는 영향들을 평가하였다. 또한, 전산 유체 역학과 유체-구조 연성 해석 결과를 실험적으로 검증하기 위해, 실제 뇌동맥 형상을 모사하는 인공 혈관과 심장의 맥동 유동을 구현하는 유체 순환 시스템을 설계하였다. 최종적으로 전산 유체 역학과 유체-구조 연성 해석 결과가 실험 결과와 잘 일치하는 것을 확인하였다. 전산 혈류 해석의 기준을 수립한 후, 혈류 역학적 인자가 동맥류에 어떠한 영향을 끼치는지를 평가하기 위해 총 58개의 Paraclinoid aneurysms을 가진 뇌동맥들에 대해 전산 유체 역학과 유체-구조 연성 해석을 수행하였다. 그 결과, 뇌동맥류가 발생한 동맥의 경우 혈류 역학적 인자가 높게 발생하는 위치와 동맥류가 발생하는 위치가 잘 일치했으며, 높은 혈류 역학적 인자 발생 위치와 동맥류 발생 위치 간에는 상당한 상관관계가 있었다. 그러나 동맥류가 발생하지 않은 동맥의 경우 유의미한 관계가 없었다. 이러한 결과를 통해 혈류 역학적 인자가 뇌동맥류의 발생 메커니즘과 밀접한 관련이 있음을 확인하였다. 혈류 역학적 인자와 뇌동맥류의 상관관계를 확인한 후, 뇌동맥류 파열 위험성 예측 모델을 구축하기 위해, 파열 뇌동맥류와 비파열 뇌동맥류 간의 혈류역학적 인자의 복잡한 관계를 평가할 수 있는 합성곱 신경망(Convolutional neural network)을 사용하였고, 전산 유체 역학과 유체-구조 연성 해석을 이용하여 총 123개의 뇌동맥류의 벽면 전단 응력 (Wall shear stress)과 변형률 (Strain)의 혈류 역학적 인자들을 계산한 다음, 합성곱 신경망 학습을 위해 이들을 이미지로 변환하였다. 또한, 새로운 데이터 증강 방법을 고안하여 123개의 뇌동맥류로부터 53,136개의 이미지를 생성한 뒤, 합성곱 신경망을 학습시키고 평가하는 데 사용하였다. 합성곱 신경망 기반 파열 위험성 예측 모델에 대해 수신자 조작 특성 곡선 (Receiver operating characteristic curve, ROC)을 활용하여 예측 모델의 성능을 평가하였을 때, 가장 높은 예측 성능을 가지는 모델의 수신자 조작 특성 곡선의 아래 면적 (Area under the ROC, AUROC)은 0.883 이었다. 또한 질량 관성 모멘트에 기반한 새로운 형상학적 인자를 고안하여 파열 위험성 예측 모델의 정확성과 실용성을 개선하였다. 기존 형상학적 인자는 뇌동맥류 형상의 불규칙도를 정량적으로 평가하는데 한계가 있지만, 질량 관성 모멘트에 기반한 새로운 형상학적 인자를 활용하면 이를 효과적으로 평가할 수 있었다. 또한, 질량 관성 모멘트를 동맥류와 모혈관의 크기를 기반으로 무차원화시켜 다양한 뇌동맥류의 상대적인 형상학적 차이를 효과적으로 평가할 수 있었다. 총 125개의 뇌동맥류 (비파열 뇌동맥류 80개, 파열 뇌동맥류 45개)에 대해 혈류역학적 인자와 질량 관성 모멘트를 계산하고 각 인자들로 학습된 인공신경망 (Artificial neural network)을 구축하여 파열 위험성 예측에 활용하였다. 질량 관성 모멘트와 혈류역학적 인자를 모두 사용하여 인공 신경망 기반 파열 위험성 예측 모델을 학습 및 평가하였을 때 가장 높은 예측 성능을 보여주었다. (민감도, 85.2%, 특이도, 92.9%, AUROC, 0.941). 본 논문에서 제시한 뇌동맥류 파열 위험성 예측 모델은 여러 한계점이 있음에도 불구하고 혈류역학적 및 형상학적 인자들을 활용하여 뇌동맥류 파열 위험성을 평가할 수 있다는 것을 보여주었다. 따라서 현재 사용하고 있는 임상학적 인자들과 함께, 본 논문에서 제안한 혈류역학적 인자와 질량 관성 모멘트에 기반한 새로운 형상학적 인자는 뇌동맥류 파열 위험성을 평가하는 데 중요한 보강 인자들로 작용할 수 있다. 다만, 본 연구가 가지고 있는 한계점과 제한점들을 해결하기 위해 추가적인 연구가 필요하며, 이를 기반으로 본 논문에서 고안된 뇌동맥류 파열 위험성 평가 방법을 개선 및 보완하는 것이 필요하다.