Predicting Long-term Outcome of Intracranial Aneurysms Treated with Flow Diverters using CFD and Machine Learning
MetadataShow full item record
An intracranial aneurysm (IA) is a weakened bulging area in the arteries in the circle of Willis, caused due to their destructive remodeling of the blood vessel. About 1 in 50 Americans harbors an IA, but most aneurysms are largely asymptomatic. Rupture of an IA is a devastating event, which leads to subarachnoid hemorrhage that carries high morbidity and mortality rates. To avoid rupture of an aneurysm, two treatment strategies are available: surgical clipping and endovascular intervention. In surgical treatment, the site of the aneurysm is accessed through the skull of the patient, and a clip is placed around the neck of the aneurysm to remove it from the arterial circulation. Although it instantaneously excludes the IA, skull-based surgery is highly invasive, leading to high risks of complications. Alternatively, endovascular therapies have emerged in the last 3 decades that aim to exclude the IA from circulation by enabling thrombotic occlusion. Traditionally, endovascular coils, made of platinum wires, were deployed inside the IA to enable clot formation and thrombosis of the IA sac. However, coil embolization has a high recanalization rate, and is not suitable for some challenging aneurysms like wide necked aneurysms, giant aneurysms or fusiform type aneurysms.To treat traditionally challenging aneurysms, a new paradigm called flow diversion using highly dense metallic-stents called flow diverters (FDs) was approved by the FDA in 2011. Since its approval, FDs have become a mainstay of endovascular intervention of not only the traditionally challenging ones, but IAs of all types. Composed of 48 braided wires primarily made of cobalt-chromium-nickel alloy and platinum markers, FDs are deployed across the parent artery to enable flow stasis inside the IA sac, thus leading to thrombosis and eventual occlusion of the IA. However, despite recent success, clinical reports suggest that FDs fail to heal ~25% of aneurysms after 6 months of treatment. These unsuccessfully FD-treated aneurysms expose the patients at persistent risks of aneurysm rupture and other thromboembolic complications. Therefore, it is important to identify aneurysms that might not heal after FD-treatment to minimize complications and improve treatment outcomes. Since FD’s primary mechanism of healing is by diverting the blood flow away from the aneurysm, I hypothesized that the hemodynamic modifications induced by FDs in successfully treated IAs differ from those of the unsuccessful ones. Additionally, I asked if outcomes of FD-treatment could be made based on these differences in hemodynamics, along with aneurysm geometry and FD characteristics in patient-specific aneurysms. To that end, in this dissertation, I developed and applied a virtual intervention workflow for FD-treatment in patient-specific IA models. The first goal was to develop an algorithm that could efficiently mimic the clinical deployment of FDs in patient-specific models. Therefore, I first developed a virtual stenting workflow (VSW) that could perform FD deployment in different patient-specific models within a few minutes. The VSW workflow was based on simplifying the complex FD wires into a generic rectangular simplex mesh. This simplex mesh was then placed inside the parent vessel, and an artificial mathematical force was applied to the mesh that attracts it toward the wall of the parent vessel. During this expansion, I also incorporated a novel collision detection algorithm that kept the simplex mesh inside the vessel at all times. Once the mesh was deployed, FD wire structure was mapped on the simplex mesh and swept into 3D wires to baton final deployed FD geometry. The time taken to deploy the FD in 2 patient-specific IAs was within a minute. I also performed post FD-treatment computational fluid dynamic (CFD) simulations to establish a workflow to both deploy FDs and obtain post-treatment hemodynamics for patient-specific cases. Next chapter involves development of a novel validation methodology for CFD solvers to obtain accurate hemodynamics in patient-specific IA models. In this chapter, I performed 2D particle image velocimetry (PIV) to obtain experimental flow velocity measurements and validated the hemodynamics of our CFD solver STAR-CCM+. During the validation process, I developed a novel methodology that identifies all the sources of errors, and identifies the error of the solver as the model error, which is isolated to represent the accuracy of the CFD solver itself. Based on this validation methodology, the model error in STAR-CCM+ solver was 5.63%±5.49% on a representative line on a patient-specific IA model. In the third chapter, I investigated if any morphological parameters are associated with the clinical outcome of FD-treated IAs. I defined novel morphometric: ostium ratio (OsR) that quantified the area of the aneurysmal ostium to the remaining circumferential parent artery. Physically, OsR quantifies the fraction of the FD that would be exposed to the aneurysm. I also defined a clinically relevant 2D surrogate called the neck ratio (NR), defined as clinical neck diameter divided by average parent artery diameter. I calculated these parameters for 63 FD-treated IA patients at our center, out of which 46 were healed (occluded group) while 17 did not occlude (residual group) 6 months after FD-treatment. Among other common aneurysmal morphometric like size, neck dimeter, I calculated OsR and NR for these two groups. Comparison results showed that only OsR and NR were significantly different between occluded and residual groups. Furthermore, ROC analysis showed that OsR was a better classifier than NR with an AUC of 0.912, as compared to 0.707 for NR. These novel morphometric are clinically important biomarkers that could identify aneurysms that are amenable to FD-treatment. Chapter four shows development of a computational analysis workflow aimed at predicting the 6-month outcome of FD-treated IAs. To achieve that, I used previously developed VSW and patient-specific CFD to obtain geometrical, FD-related and hemodynamic features for 84 FD-treated aneurysm cases at our center. I then trained 4 machine learning (ML) algorithms: logistic regression (LR), support vector machine (SVM, linear and Gaussian kernel), k-nearest neighbor and neural network (NN). The 84 cases were divided into a training and testing cohort of 64 and 20, respectively. For each case, 6 geometrical features (aneurysm size, neck diameter, OsR, NR, aspect ratio and size ratio), 2 FD-related (metal coverage rate and pore-density), 4 pre-treatment hemodynamic parameters (inflow rate, aneurysm averaged velocity, shear rate, turnover time) and 4 post-treatment values of the same hemodynamic features were calculated. All 4 ML algorithms were then trained using: (1) all 16 features and (2) using only 5 significant ones that included OsR, NR, pre-treatment inflow rate and post-treatment inflow rate and average velocity. Performance of each model was tested on 20 training cases, which showed that among all models, NN and SVM with Gaussian filter with all 16 features predicted the outcome with 90% accuracy. At the end of this study, it was concluded that using all 16 parameter provides that highest accuracy, and that NN was the best-trained algorithm among all.Now that accurate predictions could be made, the next goal was to make the predictive model clinically practical. The predictive model was highly computationally costly, with predictions of a new case taking ~24 hours. The main reason for such high computational cost was the calculations of post-treatment CFD. In this chapter I investigated the effects of different features on predictive models to assess which features are truly necessary for the models. For this study, I performed computational analysis to obtain relevant geometrical (G), FD-related (D), pre- and post-treatment hemodynamic (U* and T*, respectively) features for 105 FD-treated IAs from our center. I then trained 4 sets of NN models on 79 cases in the training cohort with following features: (1) GU*DT*, (2) GU*D, (3) GU* and (4) G. When each model was tested on 26 cases, model with GU*DT* features had the highest accuracy of 92.31%, followed by GU*D, GU* and G (accuracy of 80.77%, 80.77% and 73.08%, respectively). This study shows that including post-treatment hemodynamics is very critical to achieve high prediction accuracy. This dissertation, presents a novel computational workflow to predict the long-term outcome of FD-treated IAs. Furthermore, a feasibility study on 105 patient-cases showed that the preliminary predictive model was 92.31% accurate in 6-month outcome prediction in the independent testing cohort. This dissertation also provides evidence that it is essential to perform post-treatment CFD simulations for accurate outcome predictions. Future studies should aim at exploring two potential directions: (1) make the workflow more efficient and (2) validate the robustness of predictive models. To achieve efficiency in the workflow, explicit FD-treated CFD simulations could be replaced by implicit porous media approximations that will drastically decrease the computational efficiency of the workflow. Additionally, pulsatile simulations could also be replaced by more efficient steady-state simulations, if similar predictive accuracy could be achieved. Second, the predictive models need to be validated on a larger cohort to test for their robustness to different IA locations, and patient-data from different medical centers. In conclusion, this workflow has the potential to aid the clinicians in a priori prediction of FD-treatment that will help them test different treatment strategies and optimize the treatment accordingly, which could lead to better clinical outcomes of IA patients.