System and method for autonomous deep learning morphology analysis

Abstract

Embodiments perform autonomous deep learning morphology analysis. At least one image, e.g., a medical image, associated with a patient is obtained. The obtained at least one image is processed using an autonomous deep learning model to extract at least one morphological parameter associated with a suspected pathology of the patient. Based on the at least one morphological parameter, an output analysis is generated. Computer-automated detection of and morphology characterization of aneurysms are provided in embodiments.

Description

1440.2081001 SYSTEM AND METHOD FOR AUTONOMOUS DEEP LEARNING MORPHOLOGY ANALYSIS RELATED APPLICATION

[0001] This application claims the benefit of U.S. Provisional Application No. 63 / 700,124, filed on September 27, 2024. The entire teachings of the above application are incorporated herein by reference. BACKGROUND

[0002] Interest in using artificial intelligence (AI) and machine learning (ML) models for detecting and analyzing pathologies, e.g., aneurysms, has grown over time. SUMMARY

[0003] Accurate and efficient diagnosis of intracranial aneurysms (IAs) can enable properly managing and preventing the potentially deleterious consequences of aneurysm ruptures. Three-dimensional (3D) aneurysm morphology characteristics have the most potential to aid in predicting aneurysm rupture. With conventional approaches, however, 3D aneurysm morphology characteristics can only be obtained via manually annotating a brain scan (images) of a patient. This limitation of existing approaches prevents integrating 3D morphology characteristics for use in the clinic. Although some traditional approaches have automated IA detection with varying levels of success, none have been able to automate 3D aneurysm morphology characteristic extraction. Thus, functionality with improved accuracy, efficiency, and flexibility is needed. Embodiments provide such functionality addressing the shortcomings and disadvantages in the art. For example, embodiments deliver systems and methods for fully autonomous IA detection, intracranial vasculature segmentation, and 3D aneurysm morphological characterization. Embodiments also provide aneurysm morphological parameters for predicting aneurysm rupture.

[0004] An example embodiment is directed to a computer-implemented method of autonomous deep learning morphology analysis. The method begins by obtaining at least one image associated with a patient. Next, the obtained at least one image is processed (including segmentation, pixel clustering, and / or the like) using an autonomous deep learning model to extract at least one morphological parameter associated with a suspected pathology of the patient. Based on the extracted at least one morphological parameter, an output analysis is - 1 - 4224511.v11440.2081001 generated. The obtaining, processing, and generating are automatically performed by one or more digital processors. According to an example embodiment, the deep learning model may be a neural network model, e.g., a convolutional neural network model such as a semantic segmentation model—including a “stacked” semantic segmentation model that comprises a binary semantic segmentation model and a multiclass semantic segmentation model stacked on top of (or otherwise coupled to) each other to achieve a final output—for non-limiting examples, which may leverage U-Net or another suitable architecture known to those of skill in the art. In another example embodiment, the deep learning model may be implemented using, for non-limiting example, the Keras® and / or TensorFlow® tools for developing neural network layers, models, and / or metrics as well as for neural network training and / or inference; other known tools are also suitable. According to another example embodiment, the deep learning model may employ, for non-limiting example, SimpleITK, the simplified interface to the Insight Segmentation and Registration Toolkit (ITK) framework, for processing images input to and / or output from the model; other known interfaces and / or frameworks are also suitable. In another example embodiment, the deep learning model may utilize, for non-limiting example, the scikit-learn tool to perform clustering, e.g., of pixels or voxels; other known tools are also suitable. Further, in yet another example embodiment, the obtained at least one image may be converted into a NumPy® array format or other suitable format known to those of skill in the art—e.g., to facilitate subsequent manipulation and / or pre-processing.

[0005] In an example embodiment, the suspected pathology may be one or more aneurysms. Generating the output analysis by the one or more digital processors may include at least one of: (i) automatically detecting existence and location of the one or more aneurysms, (ii) responsively generating a 3D reconstruction of vasculature corresponding to the detected one or more aneurysms, and (iii) automatically generating a morphology characterization of the detected one or more aneurysms. According to one such example embodiment, the generated morphology characterization may include, for each detected aneurysm, indications of any combination of: volume, surface area, neck width, dome width, dome height, dome depth, proximal parent artery diameter, distal parent artery diameter, flow angle, undulation index, ellipticity index, non-sphericity index, aspect ratio, size ratio, width- parent artery diameter ratio, and height-neck ratio. In another such example embodiment, the vasculature may be intracranial vasculature. Further, according to yet another such example - 2 - 4224511.v11440.2081001 embodiment, the method may further include automatically determining, by the one or more digital processors, a rupture probability of the detected one or more aneurysms.

[0006] According to an example embodiment, the method may further include: (1) transforming, by the one or more digital processors, the generated 3D reconstruction of vasculature into a volume mesh; (2) automatically determining, by the one or more digital processors: (i) at least one boundary condition for the volume mesh, and (ii) at least one fluid property for the volume mesh; (3) responsively performing, by the one or more digital processors, a fluid dynamics simulation based on the volume mesh, the determined at least one boundary condition, and the determined at least one fluid property; and (4) automatically extracting, by the one or more digital processors, at least one hemodynamic feature based on a result of performing the fluid dynamics simulation.

[0007] In an example embodiment, transforming the generated 3D reconstruction may include, by the one or more digital processors: (1) generating a surface mesh based on the generated 3D reconstruction of vasculature and (2) transforming the generated surface mesh into the volume mesh. According to one such example embodiment, generating the surface mesh may be based on a surface smoothing technique. In another such example embodiment, the surface smoothing technique may be at least one of Laplacian surface smoothing and Taubin surface smoothing.

[0008] According to an example embodiment, the method may further include, based on one or more quality metrics of the volume mesh, automatically determining at least one of: (i) numerical stability of the volume mesh, and (ii) numerical accuracy of the volume mesh. In one such example embodiment, the one or more quality metrics may include at least one of: aspect ratio, skewness, and minimum angle.

[0009] In an example embodiment, the volume mesh may include at least one tetrahedral element that defines an interior fluid domain of the vasculature.

[0010] According to an example embodiment, the determined at least one boundary condition may include at least one of: (i) one or more inlet conditions, and (ii) one or more outlet conditions. In one such example embodiment, the method may further include automatically determining the one or more inlet conditions based on at least one of: (i) one or more patient-specific velocity profiles, and (ii) one or more standardized pulsatile waveforms. According to another such example embodiment, the method may further include automatically determining the one or more outlet conditions based on at least one of: (i) one or more fixed pressure models, and (ii) one or more resistance-based parameter models. - 3 - 4224511.v11440.2081001 Further, in yet another such example embodiment, the one or more resistance-based parameter models may include a Windkessel model.

[0011] In an example embodiment, automatically determining the at least one fluid property may be based on a Newtonian blood model or a non-Newtonian blood model.

[0012] According to an example embodiment, the determined at least one fluid property may include at least one of: density and dynamic viscosity.

[0013] In an example embodiment, the fluid dynamics simulation may be performed using a time-resolved model or a steady-state model.

[0014] According to an example embodiment, the extracted at least one hemodynamic feature may include at least one of: (i) wall shear stress (WSS), (ii) an oscillatory shear index (OSI), (iii) time-averaged WSS (TAWSS), (iv) a flow jet velocity, (v) a flow jet directionality, and (vi) intra-aneurysmal pressure.

[0015] In an example embodiment, the method may further include automatically correlating, by the one or more digital processors, the extracted at least one hemodynamic feature with the obtained at least one image.

[0016] According to an example embodiment, the method may further include automatically determining, by the one or more digital processors, vessel wall elasticity for the volume mesh. Performing the fluid dynamics simulation may include performing a fluid- structure interaction (FSI) simulation based on the determined vessel wall elasticity.

[0017] In an example embodiment, the fluid dynamics simulation may be performed by the one or more digital processors using a deep learning-based surrogate model. The at least one hemodynamic feature may be extracted in real-time of the fluid dynamics simulation.

[0018] According to an example embodiment, the method may further include automatically determining, by the one or more digital processors, a rupture probability of the detected one or more aneurysms based on the extracted at least one hemodynamic feature.

[0019] According to an example embodiment, the autonomous deep learning model may be a Brain Aneurysm Morphology (BAM) model.

[0020] In an example embodiment, the obtained at least one image may include at least one of: a computed tomography (CT) scan and a magnetic resonance imaging (MRI) scan. According to one such example embodiment, the CT scan may be a computed tomography angiography (CTA) scan having a slice thickness equal to or less than 1.5 mm. It should be noted that embodiments are not limited to scans having a slice thickness of equal to or less - 4 - 4224511.v11440.2081001 than 1.5 mm, but can also process scans having greater slice thicknesses. Other medical images, scans, slice dimensions (thicknesses), and the like are also suitable.

[0021] Another example embodiment is directed to a computer-based system for autonomous deep learning morphology analysis. The system includes a processor and a memory with computer code instructions stored thereon. The processor and the memory, with the computer code instructions, are configured to cause the system to implement any embodiments or combination of embodiments described herein.

[0022] Yet another embodiment is directed to a computer program product for autonomous deep learning morphology analysis. The computer program product includes a non-transitory computer-readable medium with computer code instructions stored thereon. The computer code instructions are configured, when executed by a processor, to cause an apparatus associated with the processor to implement any embodiments or combination of embodiments described herein.

[0023] It is noted that embodiments of the method, system, and computer program product may be configured to implement any embodiments or combination of embodiments described herein. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The foregoing will be apparent from the following more particular description of example embodiments, as illustrated in the accompanying drawings in which like reference characters refer to the same parts throughout the different views. The drawings are not necessarily to scale, emphasis instead being placed upon illustrating embodiments.

[0025] FIG.1 is a flowchart of example model training and validation procedures according to an embodiment.

[0026] FIG.2 is a block diagram of an example binary deep learning model (DLM) according to an embodiment.

[0027] FIG.3 is a block diagram of an example multiclass DLM according to an embodiment.

[0028] FIG.4 is a block diagram of an example morphology extraction model according to an embodiment.

[0029] FIG.5 illustrates example primary morphological parameters obtained by the model of FIG.4, according to an embodiment. - 5 - 4224511.v11440.2081001

[0030] FIGS.6A-6C illustrate example advanced morphology parameters obtained by the model of FIG.4, according to an embodiment.

[0031] FIGS.7A-7C are example user interfaces for an online platform according to an embodiment.

[0032] FIG.8A is a three-dimensional (3D) reconstruction of intracranial vasculature using traditional thresholding.

[0033] FIG.8B is a 3D reconstruction of the intracranial vasculature of FIG.8A using an example semantic DLM according to an embodiment.

[0034] FIG.9 illustrates an example workflow for dataset creation according to an embodiment.

[0035] FIG.10 illustrates an example Brain Aneurysm Morphology model according to an embodiment.

[0036] FIG.11 illustrates example training and validation of a binary segmentation model according to an embodiment.

[0037] FIG.12 illustrates example training and validation of a multiclass segmentation model according to an embodiment.

[0038] FIGS.13A-13D are images showing different views of an output multiclass mask according to an embodiment.

[0039] FIG.14 is a flowchart of a method for autonomous deep learning morphology analysis according to an example embodiment.

[0040] FIG.15 is a schematic view of a computer network in which embodiments may be implemented.

[0041] FIG.16 is a block diagram illustrating an example embodiment of a computer node in the computer network of FIG.15.

[0042] FIG.17 is a block diagram of an example computational cerebral fluid dynamics pipeline according to an embodiment. DETAILED DESCRIPTION

[0043] A description of example embodiments follows.

[0044] Introduction

[0045] Deep learning models (DLMs) are slowly transforming neurovascular disease diagnosis, either aiding physicians in diagnosis or suggesting diagnoses outright. This transformation has already occurred for stroke diagnosis and management, especially - 6 - 4224511.v11440.2081001 ischemic stroke. Various existing tools offer automatic detection of ischemic stroke on non- contrast computed tomography (CT) scans and magnetic resonance angiography (MRA) scans, detection of vessel occlusion on CT angiography (CTA) scans, stroke volume characterization on CT perfusion imaging, and machine learning (ML) models predicting outcomes and rehabilitation after ischemic stroke. A few providers have obtained U.S. Food and Drug Administration (FDA) approval for existing commercial tools that are used at clinics nationally and internationally. Importantly, ischemic stroke is the most prevalent type of stroke, with over 7.5 million new cases per year globally.

[0046] Subarachnoid hemorrhage (SAH), in comparison, accounts for about 500,000 new cases per year and is one of the least common types of strokes. This disparity between ischemic stroke and SAH is reflected in the lower prevalence of DLMs used to diagnose and manage IAs, which are the leading cause of SAH. Only a handful of prior studies report systems capable of IA detection on CTA and MRA. Most of these existing systems have early proof-of-concept models trained on small datasets and often without external validation. The lack of generalizability in conventional DLM performance is a factor preventing the permanent integration of existing models into the healthcare workflow. Traditional DLMs often show promising results on their training datasets for IA detection, but upon validation at external centers, the existing models have consistently failed to yield reproducible results. At a commercial level, this lack of reproducible results for conventional systems was recently observed after the first significant validation of the existing Viz™ Aneurysm DLM, which resulted in a sensitivity of 62% among eight certified stroke centers in the U.S. over seven months. Indeed, significant improvements are needed to facilitate the integration of traditional models to allow the right balance between achieving good performance on training data and ensuring reproducibility of the performance at other centers. In the pursuit of high sensitivity and specificity, existing models can be easily overfitted by overtraining on an internal dataset, which may lack diversity in diagnostic imaging brand types, using only the latest generation of CT manufacturing models. Overfitting of conventional models can even be due to lack of variability in a dose of contrast material administered, e.g., for angiography performed by the same specialists.

[0047] Compared to industry, academia has contributed more to innovation in automating IA detection. Among existing DLMs in the literature available for public validation, one research group developed the first model showing potential clinical use while having a training and validation sample size greater than FDA-cleared counterparts. The research - 7 - 4224511.v11440.2081001 group’s DLM was the first to show a moderate performance improvement upon external validation. This achievement was made possible by training on the largest dataset at the time, which included 1,177 CTAs, covering three of the four major CT brands, and externally validating at three other domestic centers covering the eastern region of China. Specifically, internal validation of the existing research group DLM resulted in an overall sensitivity and specificity of 81.0% and 86.9%, respectively. In comparison, external domestic validation yielded similar performance with a sensitivity and specificity of 76.1% and 80.2%, 73.0% and 76.1%, and 78.9% and 74.6%, respectively, at the three domestic external centers. However, this existing DLM was never integrated into an official Picture Archiving and Communications System (PACS) provider or a stand-alone platform, likely due to the limitation of requiring post-processed CTAs that have undergone skull subtraction (e.g., bone removal). All scans used for training and validating the existing research group DLM were manually post-processed with skull subtraction, which is not a feature offered by all PACS providers. Moreover, traditional skull subtraction approaches commonly fail to preserve the integrity of arteries coursing along the bone. The internal carotid artery (ICA) is especially vulnerable to this phenomenon because conventional skull subtraction often retains fragments of the petrous bone as a part of the ICA. Alternatively, sections of the ICA are sometimes subtracted from the skull by existing approaches due to being mistakenly recognized as bone, mainly when containing atherosclerosis. This is likely why a significant portion of IAs misclassified by the existing research group DLM were located along the ICA. Following the initial release of its existing DLM, the research group recently published a follow-up study with a significantly larger dataset. The follow-up study addressed the conventional skull subtraction limitation by re-training the existing research group’s DLM on raw CTAs. However, this updated existing research group model is not available to the public or for private commercial use, which prevents further validation among centers not affiliated with the research group. Despite the substantial breakthroughs reported within the research group’s follow-up study, which are indeed profound, key signs of selection bias are shown within the training and validation cohorts used, which raises questions regarding the updated existing research group DLM’s generalizability in the real world. Overall, the research group’s previous existing model using skull subtraction remains the only DLM in the literature that has demonstrated a large (e.g., >1,000 scans) and diverse training cohort while performing external validation and that provides the ability for individuals not affiliated with the research group to further validate the group’s findings, which is important for any DLM. - 8 - 4224511.v11440.2081001

[0048] Among the few existing models showing generalizable performance upon external validation are two recent existing systems that reported good sensitivity and specificity with large sample sizes. Even with perfect accuracy—which they have not attained in any case— these conventional models still cannot provide morphological information regarding an aneurysm and a parent artery. Such morphological parameters are important for accurate aneurysm rupture risk determination and stratification. Traditionally, the decision to treat an aneurysm was based on crude clinical and radiographic data. The diagnostic images were mainly utilized to extract qualitative variables, including blebbing, bifurcation / branch point localization, etc. As existing imaging analysis tools have advanced, more complex morphological parameters have been recognized as significant aneurysm rupture risk predictors. The literature in recent years has shown that these more advanced two- dimensional (2D) and three-dimensional (3D) aneurysm morphology variables consistently outperform the traditional predictors for aneurysm rupture. A recent meta-analysis found that advanced morphological parameters, including size ratio, aspect ratio, pressure loss coefficient, and wall shear stress, best predicted the risk of rupture in small aneurysms. These variables were superior to traditional clinical data and simple imaging one-dimensional (1D) quantitative variables like aneurysm dome height and diameter. An obstacle with existing systems to using such advanced aneurysm morphology parameters is that they can only be obtained upon manual annotation of a scan, requiring manual classification of each pixel slice by slice as an artery or aneurysm. The lack of autonomous capabilities in conventional systems for this process of obtaining advanced aneurysm morphology parameters is an underlying factor preventing the advanced morphology predictors from being integrated into the clinical treatment decision-making process. Even recent existing systems do not annotate the parent artery; the aneurysm dome alone also lacks volumetric congruence is conventional systems. Therefore, high-level predictors for aneurysm rupture associated with the parent artery are not obtainable by traditional systems. Existing systems also do not annotate the arterial network and thus fail to provide a complete or comprehensive morphology analysis that can be used in physicians’ treatment decision making processes and / or in an eventual computational flow dynamic analysis.

[0049] The concurrent advancement of artificial intelligence (AI) and ML techniques in medical imaging and the emergence of the advanced aneurysm morphological parameters described hereinabove call for methods and systems capable of obtaining such morphological - 9 - 4224511.v11440.2081001 variables autonomously. Embodiments deliver such methods and systems. Among other things, embodiments provide the following: a) Example techniques for developing and training semantic segmentation-based aneurysm detection model(s); b) Example techniques for developing and training semantic segmentation model(s) that capture intracranial vasculature and aneurysm(s) as separate classes; and c) Example techniques for developing morphology characterization model(s) that rapidly and automatically obtain advanced morphological characteristics of identified aneurysm(s).

[0050] According to an embodiment, an example DLM may be developed and trained for autonomous brain aneurysm detection on raw images, e.g., raw CTA scans. In another embodiment, an example model may leverage a custom asymmetric U-Net with attention gates in a decoder path to enhance the model’s focus on relevant information. An example model according to an embodiment was validated internally through a prospective cohort and externally through an international cohort with a combined state-of-the-art lesion-level (LL) sensitivity and specificity of 82.9% and 91.4%, respectively. Embodiments deliver the first model validated on an international cohort while demonstrating generalizable performance. Further, embodiments provide techniques for developing similar model(s) for segmenting intracranial vessels and / or morphological characterization of aneurysms, which has never been done before. A user-friendly web platform was designed for further public testing of an example aneurysm detection model according to an embodiment.

[0051] Traditional models for detecting and analyzing IAs are limited and struggle with generalizability and real-world performance. Embodiments overcome these and other shortcomings of existing systems. In an embodiment, an example DLM was developed to detect and analyze IAs on images, e.g., raw CTA scans. Using 2,187 example CTA scans (186a-186c (FIG.1)) from two academic centers (2004–2024), the example model was trained (160 (FIG.1)) by leveraging a 3D U-Net architecture and validated both prospectively (170a (FIG.1)) and internationally (170b (FIG.1)). The example model achieved high LL sensitivity (84%) and specificity (94%) on external international validation (170b (FIG.1)), outperforming existing FDA-cleared systems by detecting aneurysms missed by radiologists and working across different scanner brands and imaging conditions. Unlike conventional tools, among other things, embodiments do not require manual skull subtraction, have a low - 10 - 4224511.v11440.2081001 false positive rate, and are publicly available for testing via an example web-based platform. An example system according to an embodiment integrates two different models, one for automatic segmentation of intracranial vasculature and one for 3D morphological analysis of aneurysm(s) and their parent artery(ies). Embodiments provide innovative systems and methods for ML / AI-driven aneurysm detection and clinical decision support.

[0052] In an embodiment, a total of 1,587 example aneurysm-positive CTA scans (186a (FIG.1)) from 2004-2023 were used for training (160 (FIG.1)), which included 1,920 aneurysms. Prospective internal validation (170a (FIG.1)) was performed over nine months in Boston on 304 example CTAs (186b (FIG.1)), and an additional 303 example CTAs (186c (FIG.1)) from a high-volume European center were used for external validation (170b (FIG. 1)). Example exclusion criteria (101a-101c (FIG.1)) included any IA treatment and scans with slice thickness > 1.5 mm. Intracranial vessels, including any aneurysms, were segmented on each scan to obtain multiclass masks. The original scans, along with their corresponding masks, were used to train an example semantic segmentation convolutional neural network for aneurysm detection according to an embodiment (e.g., 300 (FIG.3)), followed by an example stacked model for intracranial vasculature segmentation according to an embodiment. An example morphology characterization model according to an embodiment (e.g., 400 (FIG.4)) was developed using advanced 3D techniques. The example detection model was then validated on an example internal prospective cohort and an example external cohort. It should be noted that validation with multiple external cohorts is also contemplated by embodiments.

[0053] In an embodiment, a total of 2,194 example CTA scans (186a-186c (FIG.1)) were used for training (160 (FIG.1)) and validation (170a / 170b (FIG.1)).1,587 patients and 1,920 aneurysms with a mean size of 5.3 ± 3.7 mm were included in the example training cohort (186a (FIG.1)). The mean age of the patients was 69.7 ± 14.9 years, and 1,203 patients (75.8%) were female. An example detection model according to an embodiment (e.g., 300 (FIG.3)) achieved a training Dice score of 0.88 and a validation Dice score of 0.76. Example prospective internal validation (160a (FIG.1)) on 304 example scans (186b (FIG.1)) yielded an LL sensitivity of 82.5% (95% confidence interval (CI) = 75.5 - 87.9) and specificity of 89.6 (95% CI = 84.5 - 93.2). Example external validation (170b (FIG.1)) on 303 example scans (186c (FIG.1)) demonstrated an on-par LL sensitivity and specificity of 83.5% (95% CI = 75.1 - 89.4) and 92.9 (95% CI = 88.8 - 95.6), respectively. Radiologist LL sensitivity from the external center was 84.5% (95% CI = 76.2 - 90.2), while 87.5% of the missed - 11 - 4224511.v11440.2081001 aneurysms were detected by the example model. An example segmentation model according to an embodiment achieved a training Dice score of 0.86 and a validation Dice score of 0.80.

[0054] Limitations of Existing Approaches from Literature

[0055] Only one existing research group DLM in the literature has been trained on a large cohort (>1,000) of raw CTAs while performing external validation. An updated version of the existing system, which modified the DLM to be re-trained on raw CTAs, was demonstrated by the same group (China Aneurysm AI Project Group) and is the first existing system to ever clinically prove that integrating DLMs into clinicians’ workflows yields synergistic results. Table 1 below includes characteristics of this existing updated research group system. The observations from the existing updated research group system are convincing, given that retrospective external validation was performed among six domestic Chinese centers, with a cumulative total of 1,198 scans, and prospective internal validation was performed among five other centers, with a cumulative total of 1,562 scans. The significant findings from the existing research group system were that the DLM helped to improve clinicians’ diagnostic performance from an area under the curve (AUC) of 79.5 (95% CI = 76.1–83.0) to an AUC of 87.8 (95% CI = 85.0–90.6) (p<0.0001). However, assessing the standalone performance of the existing DLM without radiologists involved was not the research group’s aim. Though the LL sensitivity was reported for the internal validation of the existing research group system, which yielded 86.6% (95% CI = 84.5–88.5), the LL sensitivity for the external validation was never disclosed, which makes it difficult to accurately evaluate the existing system’s external generalizability, given that patient-level (PL) sensitivity can be misleading if a high quantity of scans harbor multiple IAs, as seen in an example external cohort according to an embodiment. Instead, only the LL sensitivity of the external clinicians with and without DLM assistance was reported within the supplementary data for the existing research group system; however, this yielded surprising results. Among the six external centers, the average LL sensitivity of radiologists using the existing research group DLM was 81.0%, compared to 59.4% without the DLM’s assistance. These values are unexpectedly low, given that an internal center’s radiologists as well as external radiologists using a model according to an embodiment yielded a standalone LL sensitivity of 91.0% and 89.3%, respectively. It is unclear whether the clinicians from the institutions participating in the demonstration of the existing research group system had such a low diagnostic performance alone. Such aspects of the existing research group system create difficulty when attempting to determine how the existing research group DLM could potentially benefit other external centers. Selection biases - 12 - 4224511.v11440.2081001 within the training and validation cohorts for the existing research group system also appear to be significant, given that the existing system’s inclusion criteria limited CTAs to modern scanners with slice thicknesses < 0.75 mm, excluding patients with cerebrovascular pathologies other than IAs (e.g., giant arteriovenous malformations (AVMs), arteriovenous fistulas (AVFs), moyamoya disease (MMD), dissections, and even occlusions), and never specifying that fusiform IAs were included. These selection biases have the potential to undermine the existing research group model’s generalizability and exaggerate outcomes. Many patients with IAs, e.g., up to 20% in an example dataset according to an embodiment, also present with other pathologies that can serve as confounding variables, making accurate detection in such patients important. Additionally, the demographics for the existing research group system showed a greater prevalence of males compared to females who contributed to the scans, the cause of which is unclear. The above and other issues may warrant external validation among individuals not affiliated with the China Aneurysm AI Project Group. However, given that the existing research group DLM is not integrated into a web platform, available on an open-source basis, or in private commercial use, proving the generalizability of the existing DLM remains difficult.- 13 - 4224511.v11440.2081001Table 1: Characteristics of an example embodiment and existing models (trained and externally validated on >1,000 scans) in the literature and industry. NR = not reported. (*) = unclear whether fusiform aneurysms were excluded due to not being mentioned within the exclusion criteria. (**) = when comparing sensitivity and specificity among existing FDA-cleared models, refer to their exclusion criteria to account for selection bias.

[0056] Other existing DLMs trained on large cohorts but with significant caveats include a recent existing system with a hybrid DLM that used a previously developed arterial mapping model to feed a detection model with a 3D reconstruction as input to identify irregularities along vasculature corresponding to IAs. This unique approach allowed for aneurysm detection and reported clinically valuable information, including the stenosis degree and outflow angle–assisted calculations. However, despite nearly 4,500 scans in the existing system’s training data, the pseudo-external validation yielded lower performance than an example model according to an embodiment, with a LL sensitivity and specificity of 83.30% and 63.30%, respectively, for the existing system. This is likely because the existing system performed aneurysm detection via a bounding-box approach, which is well known to perform poorly on tiny lesions relative to a remainder of a scan. Additionally, the existing system’s validation was performed with scans from one of the institutions involved in the existing model’s training, which failed to be considered an external validation cohort. Another existing model also had a large training dataset and was appropriately externally validated. Still, the existing model was trained to prioritize sensitivity at the cost of specificity, with 3.6 false positives per case on external validation. Such a false positive rate in the existing model can significantly increase the time clinicians spend checking the existing model’s results and likely negate any benefit in efficiency.

[0057] Limitations of Existing Approaches from Industry

[0058] Currently, three existing DLMs are cleared by the FDA with trade names BriefCase, Viz Aneurysm, and Rapid ANRTN, each approved in 2022, 2023, and 2023, - 14 - 4224511.v11440.2081001 respectively. Table 1 above includes characteristics of these existing commercial systems. The first external validation ever conducted on any of these existing models, specifically the Viz model, was published in 2023. The results from the seven-month evaluation among eight certified stroke centers in the U.S. were surprisingly low, with a reported sensitivity of 62% for the existing Viz model. With a reported stated sensitivity of 93% and specificity of 89% in the existing Viz model’s FDA filings, the substantially lower performance on external validation can be explained by the existing system’s inclusion and exclusion criteria. All three existing commercial DLMs excluded fusiform aneurysms, saccular aneurysms < 4 mm, and those situated beyond the intracranial arteries’ A2, M2, and P1 segments. According to the Unruptured Cerebral Aneurysm Study (UCAS) of Japan, these exclusion criteria used in the existing industry systems result in the systems failing to cover over 40% of the IAs in the population. This discrepancy likely explains why a sensitivity of only about 60% was seen upon the first real-world external validation of the existing Viz model.

[0059] A recent patent application filed by Viz purports to disclose a system that receives patient images from a hospital server, runs AI models on the images to predict if there is an aneurysm, and then sends back a notification to various local / online locations if an aneurysm is detected. It appears that the existing Viz system can only detect aneurysms > 4-5 mm in size. According to the UCAS cohort, about 46% of all aneurysms are < 4 mm. The existing Viz system is thus only applicable for less than half of all aneurysms. Moreover, the existing Viz system can only give basic information about a size of an aneurysm, such as width and volume. The existing Viz system also lacks the ability to provide any information, such as segmentation information, about surrounding vasculature—which information is important for aneurysm rupture prediction. Similarly, the existing Viz system lacks the ability to provide any information about morphology characterization or treatment planning. In publications discussing the existing Viz system, there is no mention of the system’s ability to provide information about aneurysm size or volume, likely because the existing system’s accuracy is poor. Moreover, the existing Viz system is only designed for aneurysm detection at best, not morphology characterization. The patent application is also vague about the sensitivity and specificity of the existing Viz system. As discussed above, a recent publication showed the existing Viz system’s positive predictive value to be only 62%. Positive predictive value refers to the percentage of scans containing aneurysms that are accurately identified as such by a model. The existing Viz system is also limited to processing input CTA scans of < 1 mm in slice thickness. - 15 - 4224511.v11440.2081001

[0060] The limitations of the existing Viz system / model and other existing industry systems described above must be considered whenever these existing commercial models are used in a clinical setting. Generalizability remains an important consideration for successfully integrating a DLM in a real-world practice where the scanner, slice thickness, contrast bolus, and other factors cannot be readily controlled.

[0061] Moreover, the existing Viz system / model discussed above, and almost every other existing aneurysm detection model—from the literature or industry—at best provide rough estimations of aneurysm segmentation. Such rough estimations are only helpful with basic detection of whether an aneurysm is present on a scan or not. They also cannot be used to obtain any useful morphology information, especially because the rough estimations lack any segmentation of the remaining vasculature. Segmenting the remaining vasculature is important for morphology characterization. Even more recent existing systems have a pronounced lack of precision and accuracy. Such traditional systems do not segment the arteries and thus have extremely limited morphology analysis capabilities.

[0062] Example Advantages of Aneurysm Detection Model

[0063] An example aneurysm detection model (e.g., 300 (FIG.3)) according to an embodiment demonstrates notable improvements in several areas, for example in the model’s high sensitivity without sacrificing a false positive (FP) rate and in the model’s generalizability. The sensitivity and specificity of an example model according to an embodiment, especially for small aneurysms (70.2%), are comparable to or better than existing models in the literature (51%-78%). Table 2 below is an example confusion matrix for LL analysis per aneurysm size in validation cohorts according to an embodiment. Unlike existing models, an example detection model according to an embodiment does not sacrifice a higher FP rate to achieve the model’s high sensitivity. The existing FP rate in the literature is around or higher than 0.25-0.57 FP per true aneurysm. In an example validation cohort according to an embodiment, this rate was 0.09, which is much lower than the existing FP rate and is also complemented by a reasonably low false negative rate of 0.17, which is comparable to that of the existing literature’s 0.14-0.34. The above information leads to a conclusion that an example detection model according to an embodiment can detect all or nearly all aneurysms without requiring a clinician to spend extra time reviewing mistaken reports, unlike existing models.- 16 - 4224511.v11440.2081001Table 2: Example confusion matrix for LL analysis per aneurysm size in validation cohorts according to an embodiment.

[0064] An example DLM for aneurysm detection according to an embodiment was trained on a large and diverse dataset of 1,587 CTA scans, the first dataset to include scans performed over a timespan of nearly two decades (2004–2023). This example training methodology according to an embodiment, which includes scans from outdated CT scanner equipment, is unique in the literature and offers at least two example advantages. One example advantage is that, to compensate for a lower resolution and increase in artifacts from the outdated CT scanners, the example DLM achieves a global anatomical understanding for detecting IAs effectively. A further example advantage is the improvement in generalizability of the example DLM, especially among centers of low-income regions where access to the latest-generation scanners may be unfeasible. Similar to existing DLMs, all four major CT scanner manufacturing brands (General Electric®, Philips®, Siemens®, and Toshiba®) were included in the example training methodology according to an embodiment. However, the example training methodology also included scans with slice thicknesses up to 1.5 mm, thereby significantly increasing heterogeneity. Including various vascular pathologies and imaging characteristics allowed the model to generalize across different imaging environments and to detect IAs in the context of other cerebrovascular diseases, such as ischemic stroke, moyamoya disease, and most AVMs / AVFs. The similar performance of an example detection model according to an embodiment on scans with other cerebrovascular diseases compared to the entire cohort is evidence of the model’s diverse training cohort.

[0065] Moreover, there is currently an unusual trend among prior studies of DLM-based aneurysm detection, where existing models mostly used digital subtraction angiography (DSA)-verified CTA scans. The prior studies reflect a belief that having a corresponding DSA improves semantic segmentation performance for an existing DLM because specialists assigned to annotating ground truths have a more precise reference. While this belief is true, unless an existing DLM is designed to extract volumetrics of an aneurysm of interest—which no existing models have—existing DLMs built solely for detection may tend to prioritize obtaining a more generalizable model at the expense of improving volumetric congruence. - 17 - 4224511.v11440.2081001 Existing models trained only on DSA-verified CTA scans may be susceptible to selection biases, given that such scans overall contain higher-risk IAs. Because DSA is invasive and carries a risk of mortality and morbidity, it is uncommon to perform DSA on patients with IAs that are considered low risk, such as small IAs with no irregularities. Training existing models only on DSA-verified CTAs and thereby excluding small and low-risk IAs, which still make up a significant portion of all IAs, can prevent existing DLMs from detecting smaller lesions. An embodiment avoids the pitfalls of this conventional approach when designing an example training cohort. In an embodiment, among 1,587 example scans for training, 809 scans (51.0%) had a corresponding DSA compared to a prior study that used only CTAs with a corresponding DSA. According to another embodiment, training on all scans, e.g., CTAs, instead of only DSA-verified scans, enables an example detection model to learn real-world characteristics of IAs during training to provide greater generalizability.

[0066] Example Advantages of Segmentation Models

[0067] Segmenting intracranial vasculature is a much more difficult task on CTA scans than on their MRA and DSA counterparts. The reason is that, in the latter two modalities, the pixel intensity of the blood is sufficiently different from that of the background so as to allow easy segmentation by thresholding. On CTA scans, however, thresholding works much less consistently for many reasons, including variability in the amount of contrast in blood, variability in the strength of the CT beams, and susceptibility of the contrast in blood to be obscured by artifact and bone. For these reasons, a more nuanced technique than simple thresholding may be needed to segment vessels on CTAs. Engineering groups have explored a handful of existing approaches to perform this task with variable levels of success. In comparison, an embodiment may leverage deep learning to provide an elegant solution to this problem by automatically learning and utilizing an appropriate example set of filters to achieve accurate segmentation of vessels. An example model according to an embodiment was able to accurately segment out intracranial vessels with remarkable accuracy on most scans. By employing advanced techniques for segmenting intracranial vasculature, embodiments provide accurate and practical systems and methods to aid in diagnosing and treating IAs.

[0068] Further Example Advantages and Features

[0069] Embodiments provide an example publicly available web-based platform that enables intuitive and user-friendly assessment of example DLM generalizability using data from multiple external clinical centers. - 18 - 4224511.v11440.2081001

[0070] As shown by the example performance analysis in Table 3 below, a high sensitivity and specificity were observed in the isolated analysis on smaller AVMs / AVFs and low-grade MMD. Embodiments thus provide acceptable performance in situations with concurrent pathologies.Co-Findings on Scan Ischemic Stroke 71 85.9 86.7 (62.1 - 96.3) 85.7 (74.3 - 92.6) 73.9 (53.5 - 87.5) StrokeSubdural38 81.8 75.0 (30.1 - 95.4) 82.8 (65.5 - 92.4) 75.0 (30.1 - 95.4)AVM / AVF / MMD 27 81.5 87.5 (52.9 - 97.8) 78.9 (56.7 - 91.5) 88.9 (56.5 - 98.0) CT-Manufacturing Brands - - -- 19 - 4224511.v11440.2081001 Table 3: Performance validation of an example model according to an embodiment stratified by cohort, CT manufacturing brands, aneurysm size, and location. (ACoA = anterior communicating artery; ACA = anterior cerebral artery; MCA = middle cerebral artery; PCoA = posterior communicating artery; PCA = posterior cerebral; BA = basilar artery; SCA = superior cerebellar artery; VA = vertebral artery; PICA = posterior inferior cerebellar artery)

[0071] Moreover, embodiments provide systems and methods previously unknown to the literature for aneurysm detection on CTA scans, including detection of any AVMs / AVFs and MMD during validation.

[0072] An example sensitivity and specificity subgroup analysis of combined internal and external validation cohorts stratified by scanner brands identified no significant differences between scanner brands for embodiments. Thus, it is possible to have increased confidence that variations in scanner type do not significantly affect embodiments.

[0073] Moreover, an example subgroup analysis of pooled validation cohorts stratifying by slice thickness showed no significant difference in embodiments’ accuracy when comparing different slice thicknesses.

[0074] Example Methodology

[0075] Example Data Collection

[0076] FIG.1 is a flowchart of example model training and validation procedures 100 according to an embodiment. As shown in FIG.1, the procedures 100 include collecting raw data for model development / training 160, internal prospective validation 170a, and external validation 170b. The raw data 160, 170a, and 170b may be, e.g., CTA images, DSA images, and / or other suitable known image formats, for patients exhibiting or suspected of exhibiting one or more pathologies of interest and may be collected for the same or different patient populations and / or time periods for each of model development 160, internal validation 170a, and external validation 170b. For example, the raw data 160, 170a, and 170b may be head / neck CTA and / or DSA images for patients with one or more aneurysms and may include 2,542 CTA images collected from a Boston, MA location during January 2004– January 2024, 379 CTA images collected from the Boston location during September 2023– July 2024, and 325 CTA images with available DSA images collected from a Gdańsk, Poland location during January 2015–September 2024, respectively. Exclusion criteria 101a, 101b, and 101c may then be applied to the raw data 160, 170a, and 170b, respectively. The criteria 101a-101c may be to exclude, e.g., patients with certain surgical hardware in head / brain - 20 - 4224511.v11440.2081001 and / or with certain conditions such as AVMs / AVFs and / or certain manifestations of MMD; other known types of exclusion criteria are also suitable. Applying the exclusions 101a (e.g., 920 surgical hardware / 35 giant AVM / AVF / MMD), 101b (e.g., 67 surgical hardware / 8 giant AVM / AVF / MMD), and 101c (e.g., 21 surgical hardware / 1 giant AVM / AVF / MMD) to the raw data 160, 170a, and 170b may result in cohorts for model development / training 186a (e.g., 1,587 images), internal prospective validation 186b (e.g., 304 images), and external validation 186c (e.g., 303 images) respectively. In turn, a training / validation split 103 of, e.g., 80 / 20, may be applied to the model development cohort 186a; other known split proportions or ratios are also suitable. Continuing with FIG.1, applying the split 103 to the cohort 186a may produce a training dataset 105a (e.g., 1,270 images) and a validation dataset 105b (e.g., 317 images) for use in training and validating, respectively, a model under development. Validating a developed model on the internal prospective cohort 186b may generate positive 107a (e.g., 115 / 37.8%) and negative 109a (e.g., 189 / 62.2%) results, while validating the model on the external cohort 186c may generate positive 107b (e.g., 77 / 25.4%) and negative 109b (e.g., 226 / 74.6%) results.

[0077] With reference to FIG.1, in an example embodiment, institutional review board approval was obtained from participating institutions. The example data 160, 170a, and 170b was collected at each of two participating large academic centers. The development / training data 160 included CTA scans dating between January 2004 and January 2024 and their corresponding radiology report and patient clinical data, including age, sex, and the reason for the scan. The internal prospective validation cohort’s data 170a included consecutive scans between September 2023 and July 2024. All centers provided raw head or head and neck CTA scans and adhered to the following exclusion criteria 101a-101c: (a) patients who had surgical clips, coils, catheters, or other surgical hardware in the head; (b) patients with giant AVMs / AVFs or MMD with a Suzuki grade >2; or (c) significantly degraded image quality. Scans with other cerebrovascular diseases, including ischemic stroke, intracranial vessel occlusion, dissection, and less severe AVM, AVFs, and MMD, were not excluded. Scans either included an IA or were IA-negative, with the subset of IA-negative scans larger in the validation cohorts 170a and 170b to test sensitivity and specificity properly. All included scans had a slice thickness of ≤ 1.5 mm. Both saccular and fusiform aneurysms were included. The international center involved with the external validation 170b of a model according to an embodiment only provided CTA scans that had an available DSA performed within six months. - 21 - 4224511.v11440.2081001

[0078] Example Scanning Protocols

[0079] In an example embodiment, the two participating academic centers each have their own standard set of procedures regarding contrast protocols and scanning techniques. An example procedure for a CTA scan at the Boston center is as follows: a non-contrast CT head is obtained with a slice thickness of 5 mm, a pre-contrast test injection of 30 cc saline at 5 cc / s is performed, then the actual contrast injection and CTA is obtained at 1mm slice thickness, then a post-contrast saline flush of 40 cc at 5 cc / s is performed. The acquisition timing is determined using an automated bolus-tracking system with the region of interest at the descending aorta at the level of the carina. Depending on scanner specifications, the predefined threshold is either 100 or 150 Hounsfield units (HU). The injector protocol for contrast at the Boston center is weight-based, with patients under 200 lbs receiving 50 cc contrast at 5 cc / s and patients over 200 lbs receiving 70 cc at 5 cc / s. The CTA head-only scan range is from mid-C2 through the vertex, and the CTA head and neck scan range is from the carina through the vertex. The kilovolt (kV) and milliampere (mA) specifications differ depending on the scanner. At the international center, an example protocol is identical to Boston except for the volumes of the injections. A pre-contrast test injection of 20 cc saline at 5 cc / s, followed by 70 cc of contrast at 5 cc / s, and a postcontrast saline flush of 20 cc at 5 cc / s are administered.

[0080] Example Dataset Preparation

[0081] With reference to FIG.1, 2,187 example CTA scans were used to train and validate models according to an embodiment. Specifically, the example models were trained on 1,587 example internal scans 186a, split into an, e.g., 80 / 20, ratio 103 for training and validation, with 1,270 and 317 example scans in groups 105a and 105b, respectively. The external validation datasets, the international 170b and internal prospective 170a cohorts, included 303 and 304 example scans, respectively. For the prospective 170a dataset, 134 scans (44.7%) were aneurysm positive 107a, and 104 scans (34.7%) were aneurism positive 107b for the international 170b dataset.

[0082] In an embodiment, an example scan annotation guideline was established that involved a threshold range of 150-400 HU to annotate the aneurysm dome while leaving a 1 (one) voxel margin from the outermost visible border of the aneurysm to account for the “blooming effect” or “contrast blooming.” The 1 (one) voxel margin offers a 0.488-0.625 mm margin for reduction, given the dimensions standardized during resampling. For saccular aneurysms, the annotations for the aneurysm dome terminated at the neck of the aneurysm. In - 22 - 4224511.v11440.2081001 contrast, annotations for fusiform aneurysms encompassed the entire volume of the diseased vessel from the proximal to the distal end of the fusiform dilation. All scans were segmented using the 3D Slicer® tool (version 5.6.1, Slicer Community) with three example classes—one for the background, one for the intracranial vasculature, and one for aneurysms; other known segmentation tools are also suitable. The example aneurysm class included the neck and dome of the aneurysm, with the remaining parent artery lumen contained in the example vasculature class. For patients without a corresponding DSA, the Digital Imaging and Communications in Medicine (DICOM) series, original radiologist reports, clinical histories, and previous and follow-up examinations were used to establish the best possible reference for annotation. IAs confirmed on the radiologist reports were annotated as true positives, and in the case of new IAs being identified that were previously missed by the original radiologists, a senior neurosurgeon with over 45 years of experience confirmed the diagnosis. Upon completion of annotations to the training cohort 186a (FIG.1), quality control of the annotations was performed by randomly selecting 320 scans (20%) for review by the radiologists. If the initial 20% were deemed inadequate, the selected cases would be re- annotated, and an additional fraction of the training cohort 186a would be further reviewed. However, all 320 scans from the first phase of quality control were later found to be satisfactory.

[0083] Example Data Preprocessing

[0084] In an embodiment, the original example DICOM files were deidentified before any preprocessing or model development by removing patient identifiers such as name, birth date, and medical record number. DICOM files were then converted to Nearly Raw Raster Data (NRRD) format and resampled to voxel spacing of 0.488 mm x 0.488 mm x 0.688 mm, the median raw scan voxel spacing in the x, y, and z planes, respectively. After resampling, pixel intensity was clipped to exclude outlier pixels between -2,000 and 3,000 HU, then normalized on a scale from 0 (zero) to 1 (one). Every scan was then inputted into a custom skull subtraction model to identify the boundaries of the skull and crop each scan, leaving a minimum 50-pixel margin around the skull. According to an example embodiment, skull subtraction may not be performed for a given study, so as to train a model on as much information as possible. Then, the dataset was amplified by augmentation via rotation of each scan +90 (plus ninety), 0 (zero), and -90 (minus ninety) degrees along the z-axis and mirroring each scan along the mid-sagittal plane. This augmentation effectively multiplied the training dataset sixfold. - 23 - 4224511.v11440.2081001

[0085] According to an embodiment, the last example step in preprocessing is patch creation. Each example scan and its corresponding mask were patched into patches with dimensions of 128 x 128 x 128 voxels and 25% overlap. With example resampling parameters, every patch covered about 62.5 mm x 62.5 mm x 80 mm or 312.5 mm3. In an embodiment, selection of patch dimensions may involve consideration of various example factors. For instance, it may be desirable to input the largest patches possible so a model can have the most contextual information for each voxel. A prior study tested different patch sizes and found a direct correlation between patch size and a model’s performance. There may be a tradeoff between increasing patch size and increased memory requirements for every training iteration. In an embodiment, an alternative to increasing patch size may be to increase nodes in a model, making the model more complex and computationally expensive. According to another embodiment, testing indicated that a, e.g.,128 x 128 x 128, patch size was most suitable while keeping a model lightweight. In an embodiment, a, e.g., 25%, overlap may be incorporated so that any arteries or aneurysms at the edge of one patch are more centrally located in at least one other patch. Other known patch sizes and overlap values, e.g., 160x160x160 with 50% overlap, are also suitable. According to another embodiment, incorporation of overlap may serve to compensate for the phenomenon where a convolution kernel loses some edge information in an input array.

[0086] Example Binary Segmentation Model Development and Training

[0087] FIG.2 is a block diagram of an example binary DLM 200 according to an embodiment, which may be referred to interchangeably herein as a binary semantic segmentation model. As shown in FIG.2, the binary DLM 200 takes an image 268, e.g., a preprocessed CTA scan, split into one or more patch(es) 223 as input and produces as output a binary mask 272 representing a 3D reconstruction of intracranial vasculature. In an embodiment, the model 200 may employ a 3D U-Net architecture; other known architectures are also suitable. According to another embodiment, encoder 242 and decoder 244 paths of the model 200 may be asymmetric. For instance, except for input layer 242a-242b and bottom layer 242l, the encoder 242 layers may each have three convolution operations, e.g., layers 242c-242e, 242f-242h, and 242i-242k. In contrast, except for bottom layer 244a, the decoder layers 244 may have two convolution operations each, e.g., layers 244b / 244c, 244e / 244f, 244g / 244h, and 244i / 244j. The binary DLM 200 may further include an output layer 246 that processes results of the decoder layers 244i-244j and generates an output patch 225 and in turn the binary mask 272. - 24 - 4224511.v11440.2081001

[0088] In an embodiment, the output layer 246 may take high-level decoded feature maps (not shown) from the decoder blocks 224 and project the feature maps to a desired output format. According to another embodiment, an output of the layer 246 may have 1 (one) channel. In one such embodiment, this means that the model 200 produces a segmentation map (e.g., 225) having the same spatial size as the input 223, e.g., 128×128×128, with voxel- wise probabilities or labels. According to an embodiment, the activation function applied, e.g., a sigmoid activation function for binary segmentation, may convert the final convolution output(s) 246 into interpretable probabilities. In another embodiment, the output can then be thresholded or otherwise post-processed to form the predicted segmentation mask 272 (e.g., highlighting the arteries). According to an embodiment, the output layer 246 may map the learned features back to the voxel-wise prediction space, so each voxel in the input volume 223 has a corresponding prediction in the output 225.

[0089] Continuing with FIG.2, in an embodiment, attention gates 248a-248d may be added in the decoder path 244 before the respective convolution operations 244a-244j so as to use only relevant spatial information from the encoder path 242. Typically, in a standard U- Net architecture, spatial information from an encoder path (e.g., 242a-242k) is directly concatenated or added to a corresponding layer in a decoder path (e.g., 244a-244h). An embodiment may leverage the insight that not all such spatial information is relevant and that adding such information in its entirety can introduce unnecessary noise into the decoder layers 244. In an embodiment, the attention gates 248a-248d may function by filtering spatial information from skip connections 252a-252d before it is concatenated with the respective decoder layers 244b-244j so that the model 200 “pays attention” only to relevant spatial information. According to another embodiment, the example attention gate architecture 248 takes as inputs spatial information 252 from an encoder layer and results 254 of a decoder layer.

[0090] Continuing with FIG.2, an example attention gate 248 may receive a gating vector 252 containing contextual information from a skip connection 252a-252d and input features 254 produced by a decoder layer(s) 244a-244h. The features 252 and 254 may be linearly transformed 256a using channel-wise 1x1x1 convolutions 258a. In turn, a non-linear activation function 262a, e.g., a rectified linear unit (ReLU) activation function, may be used to generate attention coefficients (not shown); other known non-linear activation functions are also suitable. To continue, a further linear transformation 258b may be applied to the attention coefficients, followed by a second activation function 262b, e.g., a sigmoid or - 25 - 4224511.v11440.2081001 softmax activation function, to normalize the attention coefficients; other known activation functions are also suitable. An output 264 of the example attention gate 248 may be element- wise multiplication 256b of the input feature maps 254 and attention coefficients produced by the activation function 262b.

[0091] Continuing with FIG.2, in an embodiment, an example attention gate 248 may operate according to a standard U-Net architecture, where encoder 242 feature(s) are passed directly to the decoder 244 via skip connections 252. According to another embodiment, an example attention gate 248 may learn to suppress irrelevant background features (e.g., bone, soft tissue, noise, etc.) and highlight task-relevant structures (such as arteries or aneurysms). In this way, the decoder 244 may only receive the most informative spatial features, improving segmentation of small, hard-to-detect objects. According to an embodiment, an example attention gate 248 may act like a feature filter that causes a deep learning model (e.g., 200) such as a neural network to focus on arteries / aneurysms and disregard other structures before the skip connections 252 are fused.

[0092] The model 200 may leverage a U-Net architecture and may include, e.g., 30, convolutional layers with attention gates 248a-248d built into the decoder 244 to enhance performance by accounting for global contextual information.

[0093] Nearly all conventional architectures may require inputs with standardized dimensions. It may be challenging to utilize medical images, e.g., CT and MRA scans, in traditional architectures because such imaging modalities can have significant variability in the dimensions of the resulting scans. Additionally, full CT scans may have a large memory requirement, especially as they undergo the various operations of an existing system. Even current state-of-the-art graphics processing units (GPUs) may thus be unable to support full CT scans as inputs. The most common traditional solution among deep learning engineers for these problems of large memory requirements and lack of standardized dimensions is to break images into 3D patches or cubes. These patches (e.g., 223 (FIG.2)), which each have a fraction of a complete image, e.g., CTA, of 3D anatomy, may be fed into a model. Using such patches, however, raises a new problem: with every training step, the model may update its weights based only on a given fraction of the 3D anatomy in the complete image. For example, some patches may contain the entire Circle of Willis, while others may contain irrelevant information like the skull and outer cortex.

[0094] There are a few conventional solutions for the above problem associated with using patches. A recent popular existing solution has been using transformers, which can be - 26 - 4224511.v11440.2081001 advantageous according to prior studies. However, implementing transformers may incur substantially greater memory requirements and may often necessitate reducing an input patch size, which may detract from or negate the advantage of using transformers in the first place. An example embodiment may leverage the insight that using transformers can have significant drawbacks, and, instead of transformers, may employ a different mechanism to better understand spatial relationships between patches. In an embodiment, the model 200 (FIG.2) may use the attention gates 248a-248d (FIG.2) in the decoder path 244 (FIG.2) to address the shortcomings of conventional transformer-based solutions. According to another embodiment, the attention gates 248a-248d may take the skip connections 252a-252d (FIG. 2) from the encoder path 242 (FIG.2) as inputs and focus the model 200 on relevant contextual information in the decoder layer 244. Moreover, because the model 200 may have a lower memory requirement than traditional systems, the model 200 can employ a larger patch size. In an embodiment, the scan 268 and its corresponding mask 272 may be broken up into patches, e.g., 223 (FIG.2) and 225 (FIG.2), respectively, with dimensions of, e.g., 128 x 128 x 128, voxels, the largest reported in the literature; other known patch dimensions, such as 160x160x160 and larger dimensions, are also suitable. A prior study compared various patch sizes and found a significant positive correlation between patch size and model performance on segmentation tasks. In an embodiment, scans, e.g., 268, may be resampled using median validation cohort pixel spacings of, e.g., 0.488 mm x 0.488 mm, and slice thickness of, e.g., 0.625 mm, with patch covering, e.g., 62 mm x 62 mm x 80 mm. According to another embodiment, a median number of patches, e.g., 223, per complete scan may be 18, using a, e.g., 25%, patch overlap, while a range of patches per scan may be 16-20.

[0095] Training conventional systems to detect brain aneurysms with high sensitivity and specificity may be difficult due to class imbalance, with an aneurysm making up, e.g., <1%, of an overall scan. Even though a model (e.g., 200 (FIG.2)) may be trained on patches, it was found that in the training dataset 186a (FIG.1) of aneurysm-positive scans, only 11.1% included an aneurysm. Such class imbalances may pose significant challenges to deep learning engineers and require unique solutions. An embodiment may utilize a novel approach to aneurysm detection by screening patches for aneurysms, which may yield a total of, e.g., 3,156, example patches, and matching the aneurysm-positive patches with an additional, e.g., 3,156, randomly selected aneurysm-negative example patches, which may cover only the cerebrovascular network, to allow an example model, e.g., 200 (FIG.2), to better distinguish healthy vasculatures from pathological ones. In an embodiment, with a total - 27 - 4224511.v11440.2081001 of 6,312 example patches for training and data-augmentation techniques, an existing approach for amplifying datasets in deep learning via applying rotations and axis flipping was employed, and the dataset was further amplified 16-fold; other known techniques for amplifying datasets are also suitable. A total of 100,992 example patches (50% positive and 50% negative) were used to train an example model according to an embodiment, e.g., 200. Before reaching the example model, the example inputs were clipped to [-2000, 3000] HU to account for outlier pixels and then normalized to [0, 1]. The example model was trained to optimize Tversky loss with an alpha of 0.4 and a beta of 0.99; other known loss metrics and settings for loss metrics are also suitable. The Adaptive Moment Estimation (Adam) optimizer was used with an initial learning rate of 0.0001, which was halved if training loss did not decrease within three epochs; other known optimizers and learning rates are also suitable. A total of 50 epochs were trained with a batch size of 40; other known epoch and batch size settings are also suitable. For hardware, two NVIDIA® A100 GPUs were used, yielding 160 gigabytes (GB) of video random-access memory (VRAM) via parallel processing; other known hardware components are also suitable. The example model was trained over 48 hours to complete the 50 epochs.

[0096] Example Multiclass and Stacked Segmentation Model Development

[0097] FIG.3 is a block diagram of an example multiclass DLM 300 according to an embodiment, which may be referred to interchangeably herein as a multiclass semantic segmentation model or detection model. As shown in FIG.3, the multiclass DLM 300 takes as input a 3D array (not shown) having two concatenated channels 368 and 372. One channel of the input 3D array may be an image 368, e.g., a preprocessed CTA scan, while the other channel may be a binary mask 372 representing a 3D reconstruction of intracranial vasculature. In an embodiment, the mask 372 may be generated by the model 200 of FIG.2. Continuing with FIG.3, the multiclass DLM 300 produces as output a multiclass mask 376. According to an embodiment, the cerebral artery and aneurysm prediction 376 may be displayed alone in a 3D reconstruction format or in a visualization 374 of the cerebral artery and aneurysm prediction 376 applied to the original scan 368. According to an embodiment, in the multiclass mask 376, an aneurysm (not shown) may be one class, the rest of the intracranial vasculature (not shown) may be another class, and the background (not shown) may be the third mask or class. In an embodiment, an architecture of the multiclass DLM 300 may be the same as or similar to that of the binary DLM 200 of FIG.2, except that, for instance, the binary DLM 200 may use a sigmoid activation function 262b in an example - 28 - 4224511.v11440.2081001 attention gate architecture 248 to generate the cerebral artery reconstruction 272 as a single class, whereas the multiclass DLM 300 may use a softmax activation function 262b to generate the multiclass prediction 376, e.g., the artery and aneurysm. According to another embodiment, a rationale for having the 3D reconstruction of the vasculature 372 may be multifold. For example, the 3D reconstruction 372 may be helpful for a clinician to review the scan 368 manually if needed.

[0098] Continuing with FIG.3, in an embodiment, the model 300 may utilize a novel asymmetric 3D convoluted neural network for semantic segmentation of IAs from, e.g., raw CTA scans, to detect the presence of aneurysms.

[0099] An example stacked semantic segmentation model according to an embodiment may include a binary DLM, e.g., 200 (FIG.2), followed in series by a multiclass DLM, e.g., 300 (FIG.3). An input to the stacked segmentation model may be an image, e.g., 268 (FIG. 2). The stacked model may output a multiclass mask, e.g., 376 (FIG.3). In an embodiment, a binary segmentation model (e.g., 200 (FIG.2)) may provide the cerebral artery segmentation (e.g., 272 (FIG.2)) that is fed into an example multiclass model according to an embodiment (e.g., 300 (FIG.3)), which in turn may detect an aneurysm on the cerebral artery network. According to another embodiment, the example multiclass model may identify the aneurysm while re-mapping out the arteries using the binary model’s prediction as an adjunct in the process.

[0100] In an embodiment, a mask, e.g., 376 (FIG.3), that includes two distinct classes for an aneurysm and remaining vasculature can be utilized for aneurysm detection without requiring a separate classification or object detection model. According to another embodiment, using different classes may allow for manipulation of aneurysm volume separately from the remaining vasculature and easy delineation of an aneurysm neck. In an embodiment, such features may be prerequisites for calculating advanced morphology characteristic(s). Only an example model according to an embodiment can provide such features, unlike prior approaches described in the literature. One existing system built on DSA images can segment an aneurysm dome and neck, but the system fails to segment the parent artery. In an embodiment, delineating an aneurysm neck may be needed to determine advanced morphology characteristic(s) because calculating such characteristic(s) may require knowledge of this plane.

[0101] Example Segmentation Model Training - 29 - 4224511.v11440.2081001

[0102] In an embodiment, an example patching methodology with overlap (described hereinabove with respect to detection model development and training) yielded 170,424 example training patches from the training cohort 186a (FIG.1).150,026 of the example training patches contained and were used to train an example binary semantic segmentation model according to an embodiment, e.g., 200 (FIG.2). The example binary segmentation model was trained to optimize Tversky loss with an alpha of 0.4 and a beta of 0.85; other known loss metrics and settings for loss metrics are also suitable. The Adam optimizer was used along with an initial learning rate of 0.0001; the learning rate was halved if training loss did not decrease in 3 (three) epochs. Other known optimizers and learning rates are also suitable. A total of 50 epochs were trained; other known epoch and batch size settings are also suitable. All patches were used to train the example binary segmentation model in each epoch. For hardware, 2 (two) NVIDIA A100 GPUs were used via parallel processing; other known hardware components are also suitable.

[0103] For an example multiclass semantic segmentation model according to an embodiment, e.g., 300 (FIG.3), out of the 170,424 example patches from the training cohort 186a, 37,662 contained aneurysms, which were isolated to train the example multiclass segmentation model. In an embodiment, training only on aneurysm patches may help the example multiclass model better understand aneurysm features and may enhance convergence during training by excluding low-yield data containing dead space or scarce vasculature. The example binary segmentation model was trained to optimize Tversky loss with an alpha of 0.67 and a beta of 0.90; other known loss metrics and settings for loss metrics are also suitable. The same hardware, 2 (two) NVIDIA A100 GPUs, was used to train the example binary segmentation model over 50 epochs; other known hardware components and epoch settings are also suitable.

[0104] In an embodiment, once an example stacked semantic segmentation model has been trained, the stacked model may be deployed or executed on the external validation cohort 186c (FIG.1) using a single NVIDIA RTX A6000 GPU. According to another embodiment, example incoming scans may be preprocessed and then patched using the same example patching dimensions of 128 x 128 x 128 voxels and the same example overlap of 25% described hereinabove. In an embodiment, an example output predicted vasculature (e.g., 272 (FIG.2)) from an example binary semantic segmentation model (e.g., 200 (FIG. 2)), after being concatenated with the corresponding, e.g., CTA, example scan (e.g., 268 (FIG.2)), may be fed as patches (e.g., 223 (FIG.2)) into an example multiclass segmentation - 30 - 4224511.v11440.2081001 model (e.g., 300 (FIG.3)) following the same or a similar approach as in training. However, according to another embodiment, the example patches may be stitched back to the original, e.g., CTA, example scan dimension, resulting in a fusion of the example scan (e.g., 368 (FIG. 3)) and vasculature segmentation (e.g., 372 (FIG.3)) for input to the example multiclass segmentation model. In an embodiment, if the example scan is found to contain an aneurysm, an example morphology characterization model (e.g., the morphology model 400 described hereinbelow with respect to FIG.4) may analyze the stitched prediction (e.g., 476 (FIG.4)) and print or determine one or more morphology parameter(s) (e.g., 406, 408, 414, 415, 417, 419, 426, 428 (FIG.4), 504, 506, 508, 512, 514, 516, 518, 522, 524 (FIG.5), and 626, 628, 632, 636, 638, 613 (FIG.6C)). According to another embodiment, the example morphology model may perform morphology characterization or parameter extraction in, e.g., real-time or immediately.

[0105] Example Morphology Characterization Model Development

[0106] FIG.4 is a block diagram of an example morphology extraction or characterization model 400 according to an embodiment. As shown in FIG.4, the morphology model 400 processes a multiclass mask 476 and extracts and / or calculates one or more morphology parameters 478, e.g., as part of generating a morphology parameter report.

[0107] Referring to FIG.4, in an embodiment, after the multiclass DLM 300 (FIG.3) outputs a multiclass mask 376 (FIG.3), it can be used as the input 476 for the morphology characterization model 400. According to an embodiment, the morphology model 400 may perform one or more operations (not shown), e.g., one or more geometric-based functions, including, for instance: (i) obtaining spatial coordinates for aneurysm pixels, (ii) determining which clusters of pixels belong to different aneurysms, (iii) obtaining boundary plane(s) between aneurysm(s) and their parent artery(ies), and / or (iv) obtaining and / or calculating the one or more morphological parameters 478 for the aneurysm(s).

[0108] Continuing with FIG.4, in an embodiment, the one or more morphological parameters 478 may include the following example primary parameters: aneurysm volume 426 (e.g., aneurysm dome volume), aneurysm dome surface area (SA) 428, aneurysm neck SA 419, aneurysm neck diameter (not shown), aneurysm dome height 408 (e.g., perpendicular aneurysm dome height), maximum aneurysm dome height (not shown), aneurysm dome width 406, maximum aneurysm dome diameter (not shown), parent artery diameter or width 414, aneurysm inclination angle (not shown), and vessel angle (not shown). - 31 - 4224511.v11440.2081001

[0109] Continuing with FIG.4, in an embodiment, the primary parameters can be used to calculate the following example secondary parameters: aspect ratio 415, size ratio 417, dome width-parent artery diameter ratio (not shown), dome height-neck ratio (not shown), undulating index (UI) (not shown), non-sphericity index (NSI) (not shown), and ellipticity index (EI) (not shown).

[0110] An example system according to an embodiment that includes a stacked semantic segmentation model (e.g., binary DLM 200 (FIG.2) and multiclass DLM 300 (FIG.3)) and a morphology model (e.g., 400) may be referred to interchangeably herein as a Brain Aneurysm Morphology (BAM) system or model.

[0111] FIG.5 illustrates example primary or basic morphological parameters for aneurysm 502 obtained by the morphology model 400 of FIG.4, according to an embodiment. The primary morphological parameters may include: aneurysm neck diameter (N) 504, aneurysm dome width (D) 506, aneurysm dome height (H1) 508, maximum aneurysm dome height (H2) 512, proximal parent artery diameter (P1) 514, distal parent artery diameter (P2) 516, a derived value (P1.5) 518 of proximal parent artery diameter 514 * (times) 1.5, an aneurysm angle (θ1) 522, and a vessel angle (θ2) 524. In an embodiment, structures of intracranial vessels 582 may be shown in lighter shading (e.g., lighter gray shading), while a structure of the aneurysm 502 may be shown in darker shading (e.g., darker gray shading).

[0112] FIGS.6A-6C illustrate example advanced or secondary morphology parameters for aneurysm 602 obtained by the morphology model 400 of FIG.4, according to an embodiment. The advanced morphological parameters may include: aneurysm volume (V1) 626, aneurysm surface area (S1) 628, convex hull volume (V2) 632, convex hull surface area (S2) 634, UI 636, EI 638, and NSI 613. In an embodiment, structures of intracranial vessels 682 may be shown in lighter shading (e.g., lighter gray shading), while a structure of the aneurysm 602 may be shown in darker shading (e.g., darker gray shading). A structure 611 surrounding the aneurysm 602 is a convex hull which may be used to obtain the advanced parameters.

[0113] Referring to FIG.4, according to an embodiment, the morphology extraction model 400 may perform an example clustering, or connected component analysis, on the input multiclass mask 476 to obtain one or more aneurysm class-labeled clusters. In an embodiment, this clustering may be performed in a 3D array by analyzing, e.g., a 26- connected neighborhood, which may be referred to as example Equation 1 and defined as: - 32 - 4224511.v11440.2081001 ^^^(^^, ^^, ^^) = {(^^ᇱ, ^^ᇱ, ^^ᇱ) ∣ ^^^^^^(∣ ^^ − ^^′ ∣, ∣ ^^ − ^^′ ∣, ∣ ^^ − ^^′ ∣) ≤ 1 ^^^^^^ ^^(^^′, ^^′, ^^′) = 2}

[0114] where (^^, ^^, ^^) may represent coordinates of a voxel in 3D space and (^^ᇱ, ^^ᇱ, ^^ᇱ)represents coordinates of a neighboring voxel. In example Equation 1 above, the condition^^^^^^(∣ ^^ − ^^′ ∣, ∣ ^^ − ^^′ ∣, ∣ ^^ − ^^′ ∣) ≤ 1 may ensure that only voxels less than 1 (one) voxelaway in any dimension are included. The condition ^^(^^′, ^^′, ^^′) = 2 may ensure that onlyvoxels with aneurysm labels are included. Clusters may then be considered aneurysms if a given cluster meets two following example criteria: (i) maximum diameter is ≥ 3 (three) voxels and (ii) the cluster is ≥ 2 (two) voxels wide in all three cardinal axes. These criteria may effectively exclude all clusters smaller than 1.5 mm, on which the morphology extraction model 400 may not be trained. Once aneurysm clusters are defined, a next step may be to define an aneurysm neck plane on the multiclass mask 476. In an embodiment, the model 400 may find all boundary voxels between the aneurysm cluster(s) and thevasculature-class voxels, which may be referred to as example Equation 2 and defined as:^^^ = {(^^, ^^, ^^) ∣ ^^(^^, ^^, ^^) = 2 ^^^^^^ ∃(^^′, ^^′, ^^′) ∈ ^^ଶ^(^^, ^^, ^^), ^^(^^′, ^^′, ^^′) = 1}

[0115] where 1 and 2 may be artery and aneurysm labels, respectively. In exampleEquation 2 above, (^^, ^^, ^^) ∣ ^^(^^, ^^, ^^) = 2 may denote voxels in an aneurysm cluster. ^^ଶ^(^^, ^^, ^^) may be, e.g., the 26-connected neighborhood of voxel (^^, ^^, ^^), similar to theprevious example connected component analysis. If any neighboring voxels have arterylabels, which may be defined as condition ^^(^^′, ^^′, ^^′) = 1, the original aneurysm voxel maythen be included in the aneurysm ^^^^^^^^ boundary points.

[0116] In an embodiment, once the neck boundary points are specified, the aneurysm neck can be approximated by fitting a 3D plane to the neck boundary points using an example least-squares method, which may be referred to as example Equation 3, as follows:

[0117] where the objective may be to minimize ^^(^^, ^^, ^^), which may be the sum ofsquared differences between the 3D plane and the actual ^^^values. ^^^may be the aneurysm neck boundary points defined previously. According to an embodiment, the aneurysm neck diameter (e.g., 504 (FIG.5)) can be obtained by finding the neck boundary points furthest apart. In another embodiment, the neck boundary points furthest apart can be found by first - 33 - 4224511.v11440.2081001 using the following example Equation 4 to project each point on the fitted 3D plane definedby example Equation 3 above:^^^^^^,^ = ^^^^^ + ^^^^^ + ^^then, the Euclidean distance between each point pair may be calculated as shown in example Equation 5 below to find the points furthest apart:

[0118] where the subscripts i and j may correspond to two different points. In exampleEquation 5 above, the largest distance max ^^^,^ may equal the aneurysm neck diameter (e.g.,504 (FIG.5)). According to an embodiment, the boundary points of the aneurysm class and background class may constitute the aneurysm surface. In another embodiment, the same or similar logic to obtain the boundary points for the aneurysm ^^^^^^^^ described hereinabove with respect to example Equation 2 can also be used to obtain these boundary points of the aneurysm class and background class. According to an embodiment, aneurysm dome height (e.g., 408 (FIG.4) or 508 (FIG.5)) can be obtained by calculating the largest perpendicular line between the neck plane and the aneurysm surface, as shown in example Equation 6below:

[0119] where ^^ may include the aneurysm surface boundary points. In an embodiment, parent artery diameter (e.g., parent artery diameter 414 (FIG.4), proximal parent artery diameter 514 (FIG.5), or distal parent artery diameter 516 (FIG.5)) can be obtained similarly to example Equation 6 above by substituting ^^ with the vessel wall boundary points, ^^.

[0120] In an embodiment, to determine maximum aneurysm dome height (e.g., 512 (FIG. 5)), the centroid of the aneurysm neck plane may be obtained by averaging the coordinates ofthe neck plane (^^^, ^^^, ^^^), followed by finding the furthest point in ^^ from the centroid, asshown in example Equation 7 below:^^^^^^^^^^^^^^ ^^^^^^^^ ℎ^^^^^^ℎ^^ = m^∈aௌx ^(^^^ − ^^^)ଶ + (^^^ − ^^^)ଶ + (^^^ − ^^^)ଶ- 34 - 4224511.v11440.2081001 According to an embodiment, maximum aneurysm dome diameter can be obtained similarly to example Equation 7 above by finding the two points furthest apart in ^^, as shown in example Equation 8 below: ^^^^^^^^^^^^^^ ^^^^^^^^ ^^^^^^^^^^^^^^^^ = m ଶ ଶ ଶ^,^a∈ௌx ^(^^^ − ^^^) + (^^^ − ^^^) + (^^^ − ^^^)In an embodiment, aneurysm dome width (e.g., 406 (FIG.4) or 506 (FIG.5)) can be defined by example Equation 8 below as the largest dome diameter parallel to the aneurysm neck diameter: ଶଶ ଶ ^ ^^^^^^^^ ^^^^^^^^ℎ = max ൫^^ − ^^ ൯ + ൫^^ − ^^ ൯ + ൫^^ − ^^ ൯ ^^^^^^ℎ ^^ℎ^^^^ ^^ × ^^ = 0^ ^ ^ ^ ^ ^ ^,^ ^^^^^,^∈ௌ

[0121] where ^^ and ^^ are the vectors of the aneurysm surface points and the^,^ ^^^^aneurysm neck diameter, respectively. According to an embodiment, aneurysm volume (e.g., 426 (FIG.4) or 626 (FIG.6C)) may simply be defined as the combined volume of all aneurysm voxels in a specified cluster, while aneurysm surface area (e.g., 628 (FIG.6C)) may be defined by example Equation 9 below as the sum of the exposed aneurysm surface voxels areas:

[0122] where ^^ may be the boundary points in the aneurysm surface. In example^Equation 9 above, 1 (^^ ) may be an indicator function that is 1 (one) if the face of voxel ^^^ ^ ^in direction ^^ is exposed and 0 (zero) otherwise. According to an embodiment, ^^ may be the^ଶ ଶ area of each of the different voxel faces where ^^ = 0.238 ^^^^ , ^^ = 0.305 ^^^^ , and௫௬ ௫௭ଶ ^^ = 0.305 ^^^^ ; other known values are also suitable.௬௭

[0123] In an embodiment, to determine the vessel angle (e.g., 524 (FIG.5)) and the aneurysm inclination angle (e.g., 522 (FIG.5)), the morphology extraction model 400 may obtain the parent vessel centerline by taking the midpoint of the vessel at all cross sections. This may create a “stick figure” representation of the vasculature, which may be used to determine the flow direction and the appropriate viewing plane for these angle measurements. According to another embodiment, the proximal end of the vasculature may be determined by - 35 - 4224511.v11440.2081001 the most inferior point of the representation. In an embodiment, the aneurysm angle defined by the angle between the aneurysm neck and the maximum aneurysm dome height can beobtained by example Equation 10 below:

[0124] where ^^^and ^^^^௫^^௨^ ௗ^^^ ^^^^^௧may be the vectors for the aneurysm neck plane and maximum aneurysm dome height, respectively. In an embodiment, determining vessel angle may include additional steps. For example, once the proximal end of the vasculature is determined, a cross-section of the parent vessel may be taken at the most proximal end of the aneurysm neck. The diameter of this circle, e.g., the proximal parent artery diameter 514 (FIG.5), may be obtained (^^^). Another cross-section of the vessel, e.g., the distal parent artery diameter 516 (FIG.5), may be taken 1.5*^^^the distance, e.g., the derived value 518 (FIG.5), proximally. Then, the vector connecting the midpoints of these two cross sections (^^^^^^^௧ ^^௧^^௬) may be used to calculate the vessel angle 524 (FIG.5) asshown in example Equation 11 below:

[0125] In an embodiment, the morphology extraction model 400 may calculate other secondary or advanced morphology parameters using these primary or basic parameters.

[0126] Example Statistical Analyses

[0127] According to an embodiment, for PL classification metrics, true positive was defined as having at least one true intracranial aneurysm indicated among examinations with IAs. In contrast, true negative was defined as no aneurysm reported by a reader or reviewer among examinations having no aneurysm. Notably, a false negative was determined as a diagnosis other than the true lesion among CTA examinations with an aneurysm. Other known PL classification metrics are also suitable. In another embodiment, for LL classification metrics, which may be used for assessing the reliability of DLMs for aneurysm detection, true positive was defined as an intracranial aneurysm correctly detected among all observed IAs; other known LL classification metrics are also suitable. According to an embodiment, quantitative variables were expressed as mean ± standard deviation if normally distributed, while median and interquartile range were used when non-normally distributed data. Categorical variables were expressed as frequencies and percentages. A segmentation - 36 - 4224511.v11440.2081001 model according to an embodiment, e.g., 200 (FIG.2), may detect potential aneurysm lesions in CTA image data. In an embodiment, detection results may be expressed as a Dice score (a measure of overlap between ground truth segmentation and predicted segmentation), PL sensitivity and specificity, LL sensitivity, and accuracy to evaluate automatic segmentation performance; other known scoring measures and result types are also suitable. According to another embodiment, statistical analyses may be conducted with R (version 3.5.2, R Foundation for Statistical Computing, Vienna, Austria); other known statistical analysis tools are also suitable.

[0128] Example Online Platform

[0129] FIGS.7A-7C are example user interfaces (UIs) 700a-700c, respectively, e.g., webpages, for an online platform, e.g., a web-based platform, according to an embodiment. In an embodiment, the UIs 700a-700c may provide visual instructions to users for testing the example online platform. According to another embodiment, the example online platform may be a Health Insurance Portability and Accountability Act (HIPAA)-secure encrypted website for clinicians and researchers to use BAM. In an embodiment, an output of the example online platform may be a 3D reconstruction visualization and morphology report similar to those shown in FIGS.5 and 6A-6C, described hereinabove.

[0130] A detection model according to an embodiment, e.g., 300 (FIG.3), may be tested through an example online platform, e.g., a publicly available, secure web-based platform, using a link, e.g., www.findaneurysm.com. In an embodiment, with reference to FIGS.7A- 7C, users may perform the following example steps to upload brain scans for testing the detection model: a) A user may click on link 721 in the UI 700a to upload a brain scan, e.g., an anonymized CTA scan. Clicking the link 721 may cause the UI 700b to be displayed. b) The UI 700b may prompt the user to locate folder 788 on the user’s local computer. The folder 788 may contain one or more files 792a-792n with brain scan slices, e.g., raw CTA slices. The files 792a-792n may be in, e.g., .dcm (DICOM) or Unix Executable File format; other suitable known file formats may also be used, e.g., computer-aided design (CAD) exported (EXP) files, research software experiment logs, etc. As shown in FIG.7B, the user may click “Cancel” button 794 to terminate the upload process; “Open” button 796 - 37 - 4224511.v11440.2081001 may be disabled (as indicated by, e.g., greying out) in the UI 700b because the user has not yet selected any files for uploading. c) As shown in FIG.7C, the user may identify a set of files 798 for selection (e.g., from among the files 792a-798n in the folder 788 on the user’s local computer) that comprise, e.g., an entire CTA scan. In an embodiment, the user may select the set of files or slices 798 by pressing, e.g., the “command + A” key combination on a personal computer (PC) running an Apple® operating system (OS) or the “control + A” key combination on a PC running a Microsoft® OS; other known techniques are also suitable. d) In UI 700c, the “Open” button 796 may be activated or enabled upon the occurrence of file selection 798. The user may click the activated button 796 to confirm uploading of the selected files 798 to the online platform for testing the detection model.

[0131] In an embodiment, a testing duration on the example online platform may vary based on the number of slices present. According to another embodiment, testing waiting times may be improved by upgrading backend server hardware of the example online platform. In an example embodiment, suitable known anonymization techniques may be employed to protect individual patient scans (e.g., CTA scans) and / or patient notes used in a study and to comply with hospital regulations and / or restrictions designed to protect patient privacy.

[0132] Example Results

[0133] Example Training Scans Characteristics

[0134] In an embodiment, a total of 2,542 aneurysm-positive CTA scans were completed between January 2004 and January 2024 (160 (FIG.1)). Excluded (101a (FIG.1)) from these were 920 scans (36.2%) containing intracranial surgical hardware, including clips and coils, 5 scans (0.1%) of patients diagnosed with moderate-high Suzuki grade MMD, and 30 scans (1.2%) containing giant AVMs / AVFs. Overall, 1,587 CTAs were included in the training cohort (186a (FIG.1)), with the mean patient age being 69.7 ± 14.9 years and 1,203 patients (75.8%) being females. DSAs performed within six months from the CTA were present in 809 (51.0%) of patients. Regarding additional pathological findings on imaging, ischemic stroke was present in 301 scans (20.0%), hemorrhagic stroke in 132 scans (8.3%), subdural hematomas in 139 scans (8.8%), and rare cerebrovascular disorders, including AVMs / AVF / MMD, in 121 scans (7.6%). Multiple IAs were present in 319 (20.1%) of the - 38 - 4224511.v11440.2081001 scans, yielding a total of 1,920 IAs in the training cohort (186a (FIG.1)). The mean aneurysm size was 5.3 ± 3.7 mm. Aneurysm location included: 332 aneurysms (17.3%) at the ACoA; 47 aneurysms (2.5%) at the ACA; 595 aneurysms (31.0%) at the MCA; 687 aneurysms (35.8%) at the ICA; 111 aneurysms (5.8%) at the PCoA; 23 aneurysms (1.2%) at the PCA; and 125 aneurysms (6.5%) at the BA / SCA / VA / PICA. Table 4 below includes example patient demographics and aneurysm characteristics among training cohorts according to an embodiment.Scans w / IAs 1587 (100) 111 (37.8) 188 (31.6) Scans IAs 319 (20.1)21 (6.9)43 (7.1)69.7 ± 14.9 62.1 ± 13.8 ± 60.8 ± 15.8 1920 143 103 246 Female sex 1203 (75.8) 194 (63.8) 184 (60.7) 378 (62.3) DSA present 809 (51.0)147 (48.4)303 (100) 450 (74.1) Co- on Scan- 39 - 4224511.v11440.2081001Table 4: Example patient demographics and aneurysm characteristics among training and validation / testing cohorts.

[0135] Example Validation Scans Characteristics

[0136] In an embodiment, 379 aneurysm-positive CTA scans were completed between September 2023 and July 2024 (170a (FIG.1)). Excluded (101b (FIG.1)) from these were 67 scans (17.7%) containing intracranial surgical hardware, including clips and coils, 6 scans (1.6%) of patients diagnosed with moderate-high Suzuki grade MMD, and 8 scans (2.1%) containing giant AVMs / AVFs. Overall, 304 CTAs were included in the validation cohort (186b (FIG.1)), with the mean patient age being 62.1 ± 13.8 years, while 194 patients (63.8%) were females. DSAs were available for 147 (48.4%) of the CTAs. Ischemic stroke was present in 23 scans (7.6%), hemorrhagic stroke in 20 scans (6.6%), subdural hematomas in 8 scans (2.6%), and AVMs / AVF / MMD in 26 scans (8.6%). Multiple IAs were present in 21 (6.9%) of the scans, yielding 136 IAs in the training cohort. The mean aneurysm size was 4.7 ± 2.8 mm.

[0137] In an embodiment, 325 CTAs with corresponding DSAs were completed between January 2015 and September 2024 (170b (FIG.1)). Excluded (101c (FIG.1)) from these were 21 scans (6.5%) containing intracranial surgical hardware, including clips and coils and 1 scan (0.3%) where a giant AVM was present. Following these exclusions, 303 CTAs were included in the external validation cohort (186c (FIG.1)). Ischemic stroke was present in 23 scans (7.6%), hemorrhagic stroke in 20 scans (6.6%), subdural hematomas in 8 scans (2.6%), and AVMs / AVF / MMD in 26 scans (8.6%). Multiple IAs were present in 21 scans (6.9%), yielding 136 IAs in the training cohort. The mean aneurysm size was 5.7 ± 3.7 mm.

[0138] Table 4 above includes example patient demographics and aneurysm characteristics among validation / testing cohorts according to an embodiment.

[0139] Example Detection Model Training and Validation Results

[0140] An example detection model (e.g., 300 (FIG.3)) according to an embodiment was trained on the 1,587 example scans (186a (FIG.1)) over 50 epochs, yielding a final training - 40 - 4224511.v11440.2081001 Dice score of 0.88 and a loss of 0.13; other known training epoch values are also suitable. Internal validation yielded a PL and LL sensitivity of 84.7% (95% CI = 76.8 - 90.0) and 82.5% (95% CI = 75.5 - 93.2), respectively. The specificity was 89.6% (95% CI = 84.5 - 93.2). External validation yielded a PL and LL sensitivity of 94.8% (95% CI: 87.4 - 98.0) and 83.5% (95% CI: 75.1 - 89.4), respectively. The specificity was 92.9% (95% CI = 88.8 - 95.6). Overall, the combined validation cohort had a PL sensitivity of 88.8% (95% CI = 83.5 - 92.6), a LL sensitivity of 82.9% (95% CI = 77.7 - 87.9), and a specificity of 91.4% (95% CI = 88.3 - 93.7). Further details of stratification of the combined validation cohorts by other co- existing pathological findings, aneurysm size and locations, CT brands, and CT slice thicknesses according to an embodiment are provided in Table 3 (described hereinabove).

[0141] An example confusion matrix in Table 2 (described hereinabove) outlines the prediction rates and thus false positive and false negative rates for each different aneurysm size category according to an embodiment.

[0142] Example Radiologist to Detection Model Performance Comparison

[0143] Among 326 example IAs identified within the internal prospective validation cohort (186b (FIG.1)), 13 example IAs (2.2%) were initially missed by radiologists but later diagnosed during annotation of the scans by an example detection model (e.g., 300 (FIG.3)) according to an embodiment. These IA locations were primarily at the junction between the ICA and the PCoA in 6 cases (46.2%), the cavernous segment of the ICA in 5 cases (36.8%), and the MCA in 2 cases (15.4%). Overall, the radiologists’ LL sensitivity within the internal validation cohort was 91.0% (95% CI = 85.2 - 94.6) compared to LL sensitivity of the DLM according to an embodiment of 84.0% (95% CI = 77.2 - 89.1). However, all 12 initially missed IAs were detected by the DLM.

[0144] Among 103 IAs identified within the external validation cohort (186c (FIG.1)), the radiologists missed 16 IAs (15.5%), but these were correctly diagnosed during annotation by the DLM according to an embodiment. These IA locations were also primarily at the junction of the ICA and the PCoA in 8 cases (7.8%), at the MCA in 5 cases (4.9%), at the cavernous segment of the ICA in 2 cases (1.9%), and the ACoA in 2 cases (1.9%). The DLM according to an embodiment detected 14 out of the 16 IAs missed by the radiologists, and the overall radiologist LL sensitivity from the external center was 84.5% (95% CI = 76.2 - 90.2) compared to the DLM’s sensitivity of 83.5% (95% CI = 75.1 - 89.4).

[0145] Example Segmentation and Morphology Characterization Model Development - 41 - 4224511.v11440.2081001

[0146] In an embodiment, an example stacked segmentation model was developed using the same example scans described hereinabove, along with their ground truths, including full intracranial vessels as one example class and aneurysm(s) as a second example class. According to another embodiment, a stacked or final segmentation model may have the same or similar architecture as an example detection model, the latter, e.g., as described with respect to FIG.3. In an embodiment, unlike the example detection model, the final segmentation model may include two models in a stacked configuration, with the first model (e.g., 200 (FIG.2)) outputting a binary mask of a vasculature (e.g., 272 (FIG.2)) and the second model (e.g., 300 (FIG.3)) outputting a multiclass mask (e.g., 376 (FIG.3)) with three classes similar to the ground truth. The performance of the first model after 50 training epochs was a Dice score of 0.86 for training and a Dice score of 0.80 for validation.

[0147] FIG.8A is a 3D reconstruction 800a of intracranial vasculature 827 using traditional thresholding. FIG.8B is a 3D reconstruction 800b of the intracranial vasculature 827 of FIG.8A using an example semantic DLM (e.g., 200 (FIG.2)) according to an embodiment.

[0148] FIGS.8A and 8B compare the example semantic DLM’s performance on segmentation of the intracranial vessels 827 versus the performance of conventional manual thresholding. The most striking difference between the two outputs 800a and 800b is the lack of a clear boundary between Circle of Willis region 829 and skull base bone 831 in the output 800a.

[0149] Example Dataset Creation Workflow

[0150] FIG.9 illustrates an example workflow 900 for dataset creation according to an embodiment. As shown in FIG.9, in an embodiment, the workflow 900 may include an example collection step 910 where patient scans, e.g., CTAs, for training are obtained from an internal clinical center 984 and scans for testing / validation are obtained from two external clinical centers 933a and 933b. The collection 910 may include performing the following non-limiting example operations: a) Screening patients for eligibility; b) Obtaining input file(s), e.g., DICOM file(s); c) De-identifying patient data; and d) Resampling scans to median voxel dimensions. - 42 - 4224511.v11440.2081001

[0151] In an embodiment, the workflow 900 may further include an example segmentation step 920. The segmentation 920 may include performing binary and multiclass segmentation, e.g., via the 3D Slicer tool, on the collected scans 910.

[0152] According to an embodiment, the workflow 900 may further include an example augmentation step 930. The augmentation 930 may include amplifying the segmented 920 dataset by rotating each scan +90°, 0°, and -90° along the axial axis (i.e., z-axis) and flipping each scan along the sagittal axis. In an embodiment, the augmentation 920 may effectively multiply a training dataset by, e.g., 6x.

[0153] In an embodiment, the workflow 900 may further including an example preprocessing step 940. The preprocessing 940 may include performing the following non- limiting example operations: a) Clipping image intensity; b) Normalizing pixel intensity (e.g., min-max normalization); and c) One-hot encoding of multiclass masks.

[0154] According to an embodiment, the workflow 900 may further include an example dataset creation step 950. The step 950 may include creating patches with dimensions of, e.g., 128 x 128 x 128, voxels and overlap of, e.g., 25%.

[0155] In an embodiment, results of performing the workflow 900 may include internal dataset 986 and external dataset 935.

[0156] With reference to FIG.9, in an embodiment, example CTA scans of unruptured intracranial aneurysms (UIAs) were collected 910 at a local institution 984 and at two international institutions 933a and 933b. Example inclusion criteria for the collection 910 included scans with untreated UIA without subarachnoid hemorrhage and slice thickness =< 1.5 mm; other known inclusion criteria are also suitable. Example exclusion criteria for the collection 910 included any aneurysm treatment, subarachnoid hemorrhage, and unacceptable quality precluding manual segmentation; other known exclusion criteria are also suitable. In total, 1,940 example CTA scans met the example inclusion criteria, which scans included 2,450 example aneurysms. Most example aneurysms (90.4%) were in the anterior circulation with an average maximum diameter being 5.1 mm. The collected 910 scans were segmented 920 via 3D Slicer and augmented 930 to amplify the dataset size. In turn, the segmented 920 and augmented 930 scans were used to develop and train a DLM according to an embodiment, which was then internally and externally validated.

[0157] Example BAM Model - 43 - 4224511.v11440.2081001

[0158] FIG.10 illustrates an example BAM model 1000 according to an embodiment. As shown in FIG.10, the BAM model 1000 takes an example image 1068, e.g., a CTA head scan or head and neck scan, as input and produces one or more output(s) including: an example binary mask 1072 representing a 3D reconstruction of intracranial vasculature; an example multiclass mask 1076, with an aneurysm (not shown) as one class, the rest of the intracranial vasculature (not shown) as another class, and the background (not shown) as a third class; and / or example morphology parameter(s) 1078.

[0159] Example Binary and Multiclass Segmentation Model Training / Validation

[0160] FIG.11 illustrates example training and validation of a binary segmentation model 1100 according to an embodiment. As shown in FIG.11, in an embodiment, an example internal dataset 1186 may be used for both training 1160 and internal / external validation 1170 of the binary model 1100.

[0161] FIG.12 illustrates example training and validation of a multiclass segmentation model 1200 according to an embodiment. As shown in FIG.12, in an embodiment, an example internal dataset 1286 may be used for both training 1260 and internal / external validation 1270 of the multiclass model 1200.

[0162] With reference to FIGS.11 and 12, in an embodiment, the example scans 1186 and 1286 were successfully used to create and train the models 1100 and 1200, respectively. The models 1100 and 1200 were built and trained on Amazon® Web Services (AWS®) encrypted cloud computing resources; other known computing platforms are also suitable. Final training of the models 1100 and 1200 was performed on four NVIDIA A100 SXM GPUs, which offer a combined 320GB of VRAM, 1,000GB of RAM, and 128 virtual central processing units (vCPUs); other known hardware components are also suitable. Total training time was 216 hours.

[0163] Continuing with reference to FIGS.11 and 12, in an embodiment, the models 1100 and 1200 achieved an intersection-over-union (IOU) of 95% and an accuracy of 99% on internal validation 1170 and 1270 respectively. On external validation 1170 and 1270, the models 1100 and 1200 respectively achieved an IOU of 91% and an accuracy of 93%.

[0164] Example Morphology Results

[0165] FIGS.13A-13D are images 1300a-1300d, respectively, showing different views of an output multiclass mask (e.g., 376 (FIG.3) or 1076 (FIG.10)) according to an embodiment. - 44 - 4224511.v11440.2081001

[0166] In an embodiment, measurement results of an example model had 93% volume concordance with the ground truth. Clinically, the results of the example model were deemed excellent by senior neurosurgeons.

[0167] Overview of Example Computational Cerebral Fluid Dynamics Functionality

[0168] FIG.17 is a block diagram of an example computational cerebral fluid dynamics (CCFD) pipeline 1700 according to an embodiment. As shown in FIG.17, the pipeline 1700 may take as input a 3D reconstruction of vasculature 1772, which may be segmentations of intracranial aneurysms and the surrounding vasculature generated by, e.g., the DLM 200 (FIG.20), the BAM model 1000 (FIG.10), or the method 1400 (FIG.14)). The segmentations 1772 may include a reconstruction of the arterial network from the cervical arteries / carotids and vertebral arteries through the distal intracranial branches / vessels. In turn, the pipeline 1700 may process the segmentations 1772 to generate a surface mesh 1737 of the vascular lumen. The example surface mesh 1737 may then be transformed or converted into an example volume mesh 1739. At least one boundary condition 1741 and at least one fluid property 1743 may be automatically determined for the volume mesh 1739. In an embodiment, the at least one boundary condition 1741 may be defined at inlet and / or outlet region(s) of a vascular domain (not shown) of the volume mesh 1739. According to another embodiment, the at least one boundary condition 1741 may include, e.g., an example inlet flow condition 1741a and / or an example outlet boundary condition 1741n. Continuing with FIG.17, the pipeline 1700 may include responsively performing a fluid dynamics simulation 1780 based on the volume mesh 1739, the determined at least one boundary condition 1741, and the determined at least one fluid property 1743. In an embodiment, performing the simulation 1780 may include performing a numerical simulation to solve, e.g., Navier-Stokes equations, using a solver (not shown) that utilizes, e.g., a finite volume or finite element numerical scheme (not shown). According to another embodiment, the simulation 1780 may employ time-resolved or steady-state simulation(s) (not shown) depending on a clinical or research objective. Continuing with FIG.17, based on a result (not shown) of performing the simulation 1780, the pipeline 1700 may extract at least one hemodynamic feature 1745. In an embodiment, the at least one hemodynamic feature 1745 may be quantitative hemodynamic metric(s) including, e.g., wall shear stress (WSS), oscillatory shear index (OSI), time- averaged WSS (TAWSS), flow jet velocities and / or directionality, and / or intra-aneurysmal pressure distribution(s), etc. According to another embodiment, the output(s) 1745 may support clinical assessment of aneurysm rupture risk and / or aid in procedural planning. - 45 - 4224511.v11440.2081001

[0169] In an embodiment, an example CCFD module or pipeline (e.g., 1700 (FIG.17)) may be provided to operate in conjunction with a BAM system (described hereinabove) to perform flow simulation (e.g., 1780 (FIG.17)) and hemodynamic analysis using example 3D vascular models (e.g., 272 (FIG.2) or 1072 (FIG.10)) generated by the BAM system.

[0170] For example, in an embodiment, the BAM system may reconstruct an arterial network from the cervical carotid and vertebral arteries through the distal intracranial branches, including the A4 segment of the anterior cerebral artery, M3 segment of the middle cerebral artery, and P3 segment of the posterior cerebral artery. According to another embodiment, the CCFD pipeline may use these example segmentations (e.g., 1772 (FIG.17)) as a computational domain—including a vascular domain and an interior fluid domain—for simulating blood flow (e.g., 1780 (FIG.17)) and / or extracting hemodynamic features (e.g., 1745 (FIG.17)) relevant to aneurysm rupture risk and / or aneurysm treatment planning.

[0171] Example CCFD Input Preparation and Geometry Generation

[0172] In an embodiment, example segmentations of intracranial aneurysms and the surrounding vasculature (e.g., 272 (FIG.2) or 1072 (FIG.10)) may be generated by the BAM system using a convolutional neural network trained on cross-sectional medical imaging data, e.g., CTA, magnetic resonance imaging (MRI), or MRA scans. According to another embodiment, these example segmentations may include the cervical arteries and extend through distal vessels up to A4, M3, and P3. In an embodiment, the example segmentations may be processed using standard techniques such as Marching Cubes to generate a surface mesh (e.g., 1737 (FIG.17)) of the vascular lumen; other known processing techniques are also suitable. According to another embodiment, post-processing steps, including surface smoothing (e.g., Laplacian or Taubin smoothing), may be applied to the generated surface mesh; other known post-processing techniques and surface smoothing approaches are also suitable.

[0173] Example Volumetric Meshing

[0174] In an embodiment, an example surface mesh (e.g., 1737 (FIG.17)) may be converted into a volumetric mesh (e.g., 1739 (FIG.17)) suitable for computational fluid dynamics (CFD) simulation using any suitable known open-source or commercial meshing tools. According to another embodiment, a resulting volume mesh may include tetrahedral elements that define an interior fluid domain. In an embodiment, one or more mesh quality metric(s), such as aspect ratio, skewness, and / or minimum angle, may be monitored during the conversion process to ensure numerical stability and / or accuracy; other known mesh - 46 - 4224511.v11440.2081001 quality metrics are also suitable. According to another embodiment, the monitoring may include a real-time (screen) display or printout of the quality metric(s). In an embodiment, threshold(s) for the metric value(s) may be specified as desired to ensure stability / accuracy of the conversion process. According to another embodiment, the metric(s) may be displayed as the conversion process happens to provide transparency to a user. In an embodiment, the display may further include reference range value(s) to inform the user of how well the conversion process is performing.

[0175] Example Boundary Conditions and Fluid Properties

[0176] In an embodiment, example boundary conditions (e.g., 1741 (FIG.17)) may be defined at inlet and outlet regions of a vascular domain defined by a volume mesh. According to another embodiment, inlet flow conditions (e.g., 1741a (FIG.17)) may be derived from either patient-specific velocity profiles or standardized pulsatile waveforms; other known sources for deriving inlet flow conditions are also suitable. In an embodiment, outlet boundary conditions (e.g., 1741n (FIG.17)) may be determined by utilizing a fixed pressure model or resistance-based lumped parameter models (e.g., Windkessel models); other known model(s) for determining outlet boundary conditions are also suitable. According to another embodiment, vessel walls of the volume mesh may be modeled as rigid, and a no-slip condition may be imposed; other known approaches for modeling vessel walls and determining vessel wall conditions are also suitable. In an embodiment, blood may be modeled as a Newtonian, incompressible fluid with a density of, e.g., approximately 1060 kg / m³ and a dynamic viscosity of, e.g., 0.0035 Pa·s; other known blood modeling approaches and dynamic viscosity and density values are also suitable.

[0177] Example Numerical Simulation and Solvers

[0178] In an embodiment, the Navier-Stokes equations governing incompressible fluid motion may be solved for an example volume mesh (e.g., 1739 (FIG.17)) using finite volume or finite element numerical schemes. Suitable known solvers include, but are not limited to, OpenFOAM®, SimVascular™, and ANSYS® Fluent®. According to another embodiment, depending on a clinical or research objective, the CCFD module (e.g., 1700 (FIG.17)) may employ time-resolved or steady-state simulations for performing a CFD simulation of blood flow (e.g., 1780 (FIG.17)) based on an example volume mesh (e.g., 1739 (FIG.17)).

[0179] Example Hemodynamic Output Parameters - 47 - 4224511.v11440.2081001

[0180] In an embodiment, the CCFD module (e.g., 1700 (FIG.17)) may also provide one or more quantitative hemodynamic metric(s) (e.g., 1745 (FIG.17)) including, for non- limiting examples: a) WSS; b) OSI; c) TAWSS; d) flow jet velocity; e) flow jet directionality; and / or f) intra-aneurysmal pressure distribution(s).

[0181] According to another embodiment, the above example outputs may facilitate clinical assessment of aneurysm rupture risk and / or aid in planning of treatment procedures; other known hemodynamic metrics are also suitable.

[0182] Example System Integration and Workflow

[0183] In an embodiment, the CCFD module (e.g., 1700 (FIG.17)) may be triggered after completion of an example ML / AI-based segmentation performed by the BAM system. According to another embodiment, the CCFD module may leverage the BAM system’s complete arterial coverage from the cervical vasculature through the A4, M3, and P3 segments to define example physiologically accurate CFD boundary conditions (e.g., 1741 (FIG.17)) for a volume mesh (e.g., 1739 (FIG.17)). In an embodiment, example flow simulations (e.g., 1780 (FIG.17)) resulting from the CCFD module may be integrated into a reproducible workflow that minimizes manual preprocessing and provides consistent results across patient cases. According to another embodiment, example outputs (e.g., 1745 (FIG. 17)) of the CCFD module may be aligned with original imaging data for interpretability.

[0184] Further Example CCFD Module Features

[0185] Below are non-limiting examples of additional features of the CCFD module: a) Incorporation of non-Newtonian blood rheology models; b) Fluid-structure interaction (FSI) simulations incorporating vessel wall elasticity; and c) Deep learning-based surrogate models for real-time hemodynamic prediction.

[0186] In an embodiment, the CCFD module may be configured to support one or more non-Newtonian models of blood viscosity to more accurately reflect shear-dependent fluid behavior, for example in regions of low shear such as within aneurysms. One or more non- Newtonian models—such as the Carreau-Yasuda or Casson model—can be implemented - 48 - 4224511.v11440.2081001 within an example CFD solver by modifying the constitutive equations to allow viscosity to vary as a function of local shear rate. This enhancement enables more physiologically accurate simulations of flow dynamics, for example in tortuous or low-flow vessel segments.

[0187] According to an embodiment, the CCFD module may optionally include an FSI module to simulate interactions between pulsatile blood flow and elastic deformation of arterial walls. This may be achieved by coupling a CFD solver for blood flow with a finite element model for vessel wall mechanics. Boundary mesh interfaces enables two-way transfer of forces and displacements at each time step. This configuration allows modeling of wall motion, which in turn may influence local hemodynamic parameters such as wall shear stress and pressure gradient(s), for example in thin-walled aneurysmal regions.

[0188] In an embodiment, to reduce computational time and / or resource requirements, the CCFD module may incorporate surrogate modeling techniques using deep learning. The surrogate model(s) may be trained on datasets of previously simulated CFD outputs and may learn to predict hemodynamic variables (e.g., WSS, OSI, etc.) directly from vascular geometry. Once trained, the surrogate model(s) can generate real-time flow estimations from new 3D segmentations without running full CFD simulations. This enables near- instantaneous risk profiling and visualization in time-sensitive clinical workflows.

[0189] Example Method Embodiment

[0190] FIG.14 is a flowchart of a method 1400 for autonomous deep learning morphology analysis according to an embodiment. The method 1400 is computer- implemented and may be implemented using any computing device, e.g., a processor, or combination of computing devices known to those of skill in the art.

[0191] The method 1400 begins at step 1401 by obtaining at least one image (e.g., 268 (FIG.2) or 1068 (FIG.10)), such as a medical image, associated with a patient. Next, at step 1402, the method 1400 processes the obtained at least one image using an autonomous DLM (e.g., 200 (FIG.2), 300 (FIG.3), 400 (FIG.4), 1000 (FIG.10), 1100 (FIG.11), and / or 1200 (FIG.12)) to extract at least one morphological parameter (e.g., 406, 408, 414, 415, 417, 419, 426, 428 (FIG.4), 504, 506, 508, 512, 514, 516, 518, 522, 524 (FIG.5), or 626, 628, 632, 636, 638, 613 (FIG.6C)) associated with a suspected pathology of the patient. At step 1403, based on the at least one morphological parameter, the method 1400 then generates an output analysis.

[0192] As noted, the method 1400 is computer-implemented and, as such, the functionality and effective operations, e.g., the obtaining (1401), processing (1402), and - 49 - 4224511.v11440.2081001 generating (1403), are automatically implemented by one or more digital processors. The method 1400 can also be implemented using any computer device or combination of computing devices known in the art. Among other examples, the method 1400 can be implemented using computer(s) / device(s) 50 and / or 60 described hereinbelow in relation to FIGS.15 and 16.

[0193] In an example embodiment of the method 1400, the suspected pathology may be one or more aneurysms (e.g., 502 (FIG.5) or 602 (FIG.6A)), and generating the output analysis may include at least one of: (i) detecting existence and location of the one or more aneurysms, (ii) generating a 3D reconstruction of vasculature (e.g., 272 (FIG.2) or 1072 (FIG.10)) corresponding to the detected one or more aneurysms, and (iii) generating a morphology characterization (e.g., 478 (FIG.4) or 1078 (FIG.10)) of the detected one or more aneurysms. According to one such embodiment, the generated morphology characterization may include, for each detected aneurysm, indications of any combination of: volume (e.g., 426 (FIG.4) or 626 (FIG.6C)), surface area (e.g., 419, 428 (FIG.4) or 628, 634 (FIG.6C)), neck width (e.g., 504 (FIG.5)), dome width (e.g., 406 (FIG.4) or 506 (FIG.5)), dome height (e.g., 408 (FIG.4) or 508 (FIG.5)), dome depth, proximal parent artery diameter (e.g., 514 (FIG.5)), distal parent artery diameter (e.g., 516 (FIG.5)), flow angle, UI (e.g., 636 (FIG.6)), EI (e.g., 638 (FIG.6)), NSI (e.g., 613 (FIG.6)), aspect ratio, size ratio, width- parent artery diameter ratio, and height-neck ratio. In another such embodiment, the vasculature may be intracranial vasculature. According to yet another such embodiment, the method 1400 may further include determining a rupture probability, or respective rupture probabilities, of the detected one or more aneurysms.

[0194] According to an example embodiment, the method 1400 may further include: (1) transforming the generated 3D reconstruction of vasculature into a volume mesh (e.g., 1739 (FIG.17)); (2) automatically determining (i) at least one boundary condition (e.g., 1741 (FIG. 17)) for the volume mesh, and (ii) at least one fluid property (e.g., 1743 (FIG.17)) for the volume mesh; (3) responsively performing a fluid dynamics simulation (e.g., 1780 (FIG.17)) based on the volume mesh, the determined at least one boundary condition, and the determined at least one fluid property; and (4) automatically extracting at least one hemodynamic feature (e.g., 1745 (FIG.17)) based on a result of performing the fluid dynamics simulation.

[0195] In an example embodiment of the method 1400, transforming the generated 3D reconstruction may include: (1) generating a surface mesh (e.g., 1737 (FIG.17)) based on the - 50 - 4224511.v11440.2081001 generated 3D reconstruction of vasculature and (2) transforming the generated surface mesh into the volume mesh. According to one such example embodiment, generating the surface mesh may be based on a surface smoothing technique. In another such example embodiment, the surface smoothing technique may be at least one of Laplacian surface smoothing and Taubin surface smoothing.

[0196] According to an example embodiment, the method 1400 may further include, based on one or more quality metrics of the volume mesh, automatically determining at least one of: (i) numerical stability of the volume mesh, and (ii) numerical accuracy of the volume mesh. In one such example embodiment, the one or more quality metrics may include at least one of: aspect ratio, skewness, and minimum angle.

[0197] In an example embodiment of the method 1400, the volume mesh may include at least one tetrahedral element that defines an interior fluid domain of the vasculature.

[0198] According to an example embodiment of the method 1400, the determined at least one boundary condition may include at least one of: (i) one or more inlet conditions (e.g., 1741a (FIG.17)), and (ii) one or more outlet conditions (e.g., 1741n (FIG.17)). In one such example embodiment, the method may further include automatically determining the one or more inlet conditions based on at least one of: (i) one or more patient-specific velocity profiles, and (ii) one or more standardized pulsatile waveforms. According to another such example embodiment, the method may further include automatically determining the one or more outlet conditions based on at least one of: (i) one or more fixed pressure models, and (ii) one or more resistance-based parameter models. Further, in yet another such example embodiment, the one or more resistance-based parameter models may include a Windkessel model.

[0199] In an example embodiment of the method 1400, automatically determining the at least one fluid property may be based on a Newtonian blood model or a non-Newtonian blood model.

[0200] According to an example embodiment of the method 1400, the determined at least one fluid property may include at least one of: density and dynamic viscosity.

[0201] In an example embodiment of the method 1400, the fluid dynamics simulation may be performed using a time-resolved model or a steady-state model.

[0202] According to an example embodiment of the method 1400, the extracted at least one hemodynamic feature may include at least one of: (i) WSS, (ii) an OSI, (iii) TAWSS, (iv) a flow jet velocity, (v) a flow jet directionality, and (vi) intra-aneurysmal pressure. - 51 - 4224511.v11440.2081001

[0203] In an example embodiment, the method 1400 may further include automatically correlating the extracted at least one hemodynamic feature with the obtained at least one image.

[0204] According to an example embodiment, the method 1400 may further include automatically determining vessel wall elasticity for the volume mesh. Performing the fluid dynamics simulation may include performing an FSI simulation based on the determined vessel wall elasticity.

[0205] In an example embodiment of the method 1400, the fluid dynamics simulation may be performed using a deep learning-based surrogate model. The at least one hemodynamic feature may be extracted in real-time of the fluid dynamics simulation.

[0206] According to an example embodiment, the method 1400 may further include automatically determining, by the one or more digital processors, a rupture probability of the detected one or more aneurysms based on the extracted at least one hemodynamic feature.

[0207] According to an example embodiment of the method 1400, the autonomous DLM may be a BAM model (e.g., 1000 (FIG.10)).

[0208] In an example embodiment of the method 1400, the obtained at least one image may include at least one of: a CT scan and an MRI scan. According to one such embodiment, the CT scan may be a CTA scan having a slice thickness equal to or less than 1.5 mm. Other medical images, scans, and the like are also suitable.

[0209] Example Advantages

[0210] Embodiments provide autonomous techniques for obtaining important aneurysm morphological parameters as well as innovative ML / AI-based models for autonomously detecting UIAs and extracting or retrieving important aneurysm morphological parameters from scans, e.g., CTA scans.

[0211] Moreover, embodiments can generate a 3D reconstruction of intracranial vasculature from scans having, e.g., < 1.5 mm slice thickness.

[0212] Embodiments can also detect aneurysm(s) of size, e.g., >= 2 mm, and obtain the following non-limiting example measurement(s) for the aneurysm(s): a) Volume, surface area, neck width, dome width, dome height, dome depth, proximal / distal parent artery diameter, flow angle, undulation index, ellipticity index, non-sphericity index, aspect ratio, size ratio, width-parent artery diameter ratio, and height-neck ratio. - 52 - 4224511.v11440.2081001

[0213] Among other things, embodiments provide 3D aneurysm morphology characteristics and advanced aneurysm morphology characteristics that have been shown to consistently outperform traditional clinical variables, such as patient demographics and simple size parameters, in predicting risk of aneurysm rupture.

[0214] 3D reconstructions of vasculature, such as parent artery diameter, can be important for aneurysm treatment planning. Traditionally, healthcare providers are often required to perform invasive catheter cerebral angiograms to obtain such reconstructions, but embodiments make this unnecessary.

[0215] Embodiments provide groundbreaking systems and methods that allow cerebrovascular specialists to seamlessly integrate aneurysm morphology characteristics into their clinical decision-making processes.

[0216] Moreover, embodiments can be trained on both ruptured and unruptured aneurysms and / or with MRI or MRA scans.

[0217] Embodiments can leverage an output of an example morphology model to: (i) generate a surface and volume mesh; (ii) simulate cerebral blood flow using CFD; (iii) compute WSS, OSI, and / or intra-aneurysmal pressure; and / or (iv) generate an output analysis of flow dynamics for use in aneurysm rupture risk prediction and / or therapeutic planning.

[0218] Computer Support

[0219] FIG.15 is a schematic view of a computer network in which embodiments may be implemented. Client computer(s) / devices 50 and server computer(s) 60 provide processing, storage, and input / output (I / O) devices executing application programs and the like. Client computer(s) / device(s) 50 can also be linked through communications network 70 to other computing devices, including other client device(s) / processor(s) 50 and server computer(s) 60. The communications network 70 can be part of a remote access network, a global network (e.g., the Internet), cloud computing servers or service, a worldwide collection of computers, local area or wide area networks, and gateways that currently use respective protocols (e.g., TCP / IP, Bluetooth®, etc.) to communicate with one another. Other electronic device / computer network architectures are also suitable.

[0220] FIG.16 is a block diagram illustrating an example embodiment of a computer node (e.g., client processor(s) / device(s) 50 or server computer(s) 60) in the computer network 70 of FIG.15. Each computer node 50, 60 contains system bus 79, where a bus is a set of hardware lines used for data transfer among components of a computer or processing system. The system bus 79 is essentially a shared conduit that connects different elements of a - 53 - 4224511.v11440.2081001 computer system (e.g., processor(s), disk storage, memory, I / O ports, network ports, etc.) that enables transfer of information between the elements. Attached to the system bus 79 is an I / O devices interface 82 for connecting various input and output devices (e.g., keyboard, mouse, display(s), printer(s), speaker(s), etc.) to the computer node 50, 60. A network interface 86 allows the computer node to connect to various other devices attached to a network (e.g., the network 70 of FIG.15). A memory 90 provides volatile storage for computer software instructions 92a and data 94a used to implement embodiments of the present disclosure (e.g., the procedures 100 (FIG.1), the model 200 (FIG.2), the model 300 (FIG.3), the model 400 (FIG.4), the UIs 700a-700c (FIGS.7A-7C), the workflow 900 (FIG.9), the model 1000 (FIG.10), the model 1100 (FIG.11), the model 1200 (FIG.12), the method 1400 (FIG.14), the pipeline 1700 (FIG.17), etc.). A disk storage 95 provides non-volatile storage for the computer software instructions 92b and data 94b used to implement embodiments of the present disclosure. A processing unit 84, such as one or more CPUs, one or more GPUs, or a combination thereof, for non-limiting examples, is also attached to the system bus 79 and provides for execution of computer instructions.

[0221] In one embodiment, the processor routines 92a-92b and data 94a-94b are a computer program product (generally referenced as 92), including a non-transitory, computer readable medium (e.g., a removable storage medium such as DVD-ROM(s), CD-ROM(s), diskette(s), tape(s), etc.) that provides at least a portion of the software instructions for the disclosure system. The computer program product 92 can be installed by any suitable software installation procedure, as is well known in the art. In another embodiment, at least a portion of the software instructions may also be downloaded over a cable, communication, and / or wireless connection. In other embodiments, the disclosure programs are a computer program propagated signal product embodied on a propagated signal on a propagation medium (e.g., a radio wave, an infrared wave, a laser wave, a sound wave, or an electrical wave propagated over a global network such as the Internet, or other network(s)). Such carrier medium or signals provide at least a portion of the software instructions for the present disclosure routines / program 92.

[0222] In alternative embodiments, the propagated signal is an analog carrier wave or digital signal carried on the propagated medium. For example, the propagated signal may be a digitized signal propagated over a global network (e.g., the Internet), a telecommunications network, or other networks (such as the network 70 of FIG.15). In one embodiment, the propagated signal is a signal that is transmitted over the propagation medium over a period of - 54 - 4224511.v11440.2081001 time, such as the instructions for a software application sent in packets over a network over a period of milliseconds, seconds, minutes, or longer. In another embodiment, the computer readable medium of the computer program product 92 is a propagation medium that the computer system 50 may receive and read, such as by receiving the propagation medium and identifying a propagated signal embodied in the propagation medium, as described above for computer program propagated signal product.

[0223] Generally speaking, the term “carrier medium” or transient carrier encompasses the foregoing transient signals, propagated signals, propagated medium, storage medium, and the like.

[0224] In other embodiments, the program product 92 may be implemented as a so-called Software as a Service (SaaS), or other installation or communication supporting end-users.

[0225] Embodiments or aspects thereof may be implemented in the form of hardware including but not limited to hardware circuitry, firmware, or software. If implemented in software, the software may be stored on any non-transient computer readable medium that is configured to enable a processor to load the software or subsets of instructions thereof. The processor then executes the instructions and is configured to operate or cause an apparatus to operate in a manner as described herein.

[0226] Further, hardware, firmware, software, routines, or instructions may be described herein as performing certain actions and / or functions of the data processors. However, it should be appreciated that such descriptions contained herein are merely for convenience and that such actions in fact result from computing devices, processors, controllers, or other devices executing the firmware, software, routines, instructions, etc.

[0227] It should be understood that the flow diagrams, block diagrams, and network diagrams may include more or fewer elements, be arranged differently, or be represented differently. But it further should be understood that certain implementations may dictate the block and network diagrams and the number of block and network diagrams illustrating the execution of the embodiments be implemented in a particular way.

[0228] Accordingly, further embodiments may also be implemented in a variety of computer architectures, physical, virtual, cloud computers, and / or some combination thereof, and, thus, the data processors described herein are intended for purposes of illustration only and not as a limitation of the embodiments.

[0229] The teachings of all patents, published applications, and references cited herein are incorporated by reference in their entirety. - 55 - 4224511.v11440.2081001

[0230] While example embodiments have been particularly shown and described, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the scope of the embodiments encompassed by the appended claims.

[0231] References

[0232] Miller LR, Frank J, Lee JD, et al. Abstract TP10: Artificial Intelligence Algorithm as a Diagnostic Tool for Aneurysms. Stroke.2024;55(Suppl_1):ATP10-ATP10. doi:10.1161 / str.55.suppl_1.TP10

[0233] Johnson MD, Palmisciano P, Yamani AS, Hoz SS, Prestigiacomo CJ. A Systematic Review and Meta-Analysis of 3-Dimensional Morphometric Parameters for Cerebral Aneurysms. World Neurosurgery.2024;183:214-226.e5. doi:10.1016 / j.wneu.2023.12.131

[0234] Liu Y, Wen Z, Wang Y, Zhong Y, Wang J, Hu Y, Zhou P, Guo S. Artificial intelligence in ischemic stroke images: current applications and future directions. Front Neurol.2024 Jul 10;15:1418060. doi: 10.3389 / fneur.2024.1418060. PMID: 39050128; PMCID: PMC11266078.

[0235] Luo J, Dai P, He Z, Huang Z, Liao S, Liu K. Deep learning models for ischemic stroke lesion segmentation in medical images: A survey. Comput Biol Med.2024 Jun;175:108509. doi: 10.1016 / j.compbiomed.2024.108509. Epub 2024 Apr 25. PMID: 38677171.

[0236] Bojsen JA, Elhakim MT, Graumann O, Gaist D, Nielsen M, Harbo FSG, Krag CH, Sagar MV, Kruuse C, Boesen MP, Rasmussen BSB. Artificial intelligence for MRI stroke detection: a systematic review and meta-analysis. Insights Imaging.2024 Jun 24;15(1):160. doi: 10.1186 / s13244-024-01723-7. PMID: 38913106; PMCID: PMC11196541.

[0237] Zhao Z, Zhang Y, Su J, Yang L, Pang L, Gao Y, Wang H. A comprehensive review for artificial intelligence on neuroimaging in rehabilitation of ischemic stroke. Front Neurol.2024 Mar 28;15:1367854. doi: 10.3389 / fneur.2024.1367854. PMID: 38606275; PMCID: PMC11007047.

[0238] Martin SS, Aday AW, Almarzooq ZI, Anderson CAM, Arora P, Avery CL, Baker-Smith CM, Barone Gibbs B, Beaton AZ, Boehme AK, Commodore-Mensah Y, Currie ME, Elkind MSV, Evenson KR, Generoso G, Heard DG, Hiremath S, Johansen MC, Kalani R, Kazi DS, Ko D, Liu J, Magnani JW, Michos ED, Mussolino ME, Navaneethan SD, Parikh NI, Perman SM, Poudel R, Rezk-Hanna M, Roth GA, Shah NS, St-Onge MP, Thacker EL, - 56 - 4224511.v11440.2081001 Tsao CW, Urbut SM, Van Spall HGC, Voeks JH, Wang NY, Wong ND, Wong SS, Yaffe K, Palaniappan LP; American Heart Association Council on Epidemiology and Prevention Statistics Committee and Stroke Statistics Subcommittee.2024 Heart Disease and Stroke Statistics: A Report of US and Global Data From the American Heart Association. Circulation.2024 Feb 20;149(8):e347-e913. doi: 10.1161 / CIR.0000000000001209. Epub 2024 Jan 24. Erratum in: Circulation.2024 May 7;149(19):e1164. doi: 10.1161 / CIR.0000000000001247. PMID: 38264914.6.

[0239] Zhou Z, Jin Y, Ye H, Zhang X, Liu J, Zhang W. Classification, detection, and segmentation performance of image-based AI in intracranial aneurysm: a systematic review. BMC Med Imaging.2024 Jul 2;24(1):164. doi: 10.1186 / s12880-024-01347-9. PMID: 38956538; PMCID: PMC11218239.

[0240] Colasurdo M, Shalev D, Robledo A, Vasandani V, Luna ZA, Rao AS, Garcia R, Edhayan G, Srinivasan VM, Sheth SA, Donner Y, Bibas O, Limzider N, Shaltoni H, Kan P. Validation of an automated machine learning algorithm for the detection and analysis of cerebral aneurysms. J Neurosurg.2023 Mar 3;139(4):1002-1009. doi: 10.3171 / 2023.1.JNS222304. PMID: 36883646.

[0241] Heit JJ, Honce JM, Yedavalli VS, Baccin CE, Tatit RT, Copeland K, Timpone VM. RAPID Aneurysm: Artificial intelligence for unruptured cerebral aneurysm detection on CT angiography. J Stroke Cerebrovasc Dis.2022 Oct;31(10):106690. doi: 10.1016 / j.jstrokecerebrovasdis.2022.106690. Epub 2022 Aug 4. PMID: 35933764.

[0242] Shi Z, Miao C, Schoepf UJ, Savage RH, Dargis DM, Pan C, Chai X, Li XL, Xia S, Zhang X, Gu Y, Zhang Y, Hu B, Xu W, Zhou C, Luo S, Wang H, Mao L, Liang K, Wen L, Zhou L, Yu Y, Lu GM, Zhang LJ. A clinically applicable deep-learning model for detecting intracranial aneurysm in computed tomography angiography images. Nat Commun. 2020 Nov 30;11(1):6090. doi: 10.1038 / s41467-020-19527-w. PMID: 33257700; PMCID: PMC7705757.

[0243] Bello HR, Graves JA, Rohatgi S, Vakil M, McCarty J, Van Hemert RL, Geppert S, Peterson RB. Skull Base-related Lesions at Routine Head CT from the Emergency Department: Pearls, Pitfalls, and Lessons Learned. Radiographics.2019 Jul-Aug;39(4):1161- 1182. doi: 10.1148 / rg.2019180118. PMID: 31283455.

[0244] Hu B, Shi Z, Lu L, Miao Z, Wang H, Zhou Z, Zhang F, Wang R, Luo X, Xu F, Li S, Fang X, Wang X, Yan G, Lv F, Zhang M, Sun Q, Cui G, Liu Y, Zhang S, Pan C, Hou Z, Liang H, Pan Y, Chen X, Li X, Zhou F, Schoepf UJ, Varga-Szemes A, Garrison Moore W, - 57 - 4224511.v11440.2081001 Yu Y, Hu C, Zhang LJ; China Aneurysm AI Project Group. A deep-learning model for intracranial aneurysm detection on CT angiography images in China: a stepwise, multicentre, early-stage clinical validation study. Lancet Digit Health.2024 Apr;6(4):e261-e271. doi: 10.1016 / S2589-7500(23)00268-6. PMID: 38519154.

[0245] Wei J, Song X, Wei X, Yang Z, Dai L, Wang M, Sun Z, Jin Y, Ma C, Hu C, Xie X, Yang Z, Zhang Y, Lv F, Lu J, Zhu Y, Li Y. Knowledge-Augmented Deep Learning for Segmenting and Detecting Cerebral Aneurysms With CT Angiography: A Multicenter Study. Radiology.2024 Aug;312(2):e233197. doi: 10.1148 / radiol.233197. PMID: 39162636.

[0246] Bijlenga P, Gondar R, Schilling S, Morel S, Hirsch S, Cuony J, Corniola MV, Perren F, Rüfenacht D, Schaller K. PHASES Score for the Management of Intracranial Aneurysm: A Cross-Sectional Population-Based Retrospective Study. Stroke.2017 Aug;48(8):2105-2112. doi: 10.1161 / STROKEAHA.117.017391. Epub 2017 Jun 30. PMID: 28667020.

[0247] Backes D, Rinkel GJE, Greving JP, Velthuis BK, Murayama Y, Takao H, Ishibashi T, Igase M, terBrugge KG, Agid R, Jääskeläinen JE, Lindgren AE, Koivisto T, von Und Zu Fraunberg M, Matsubara S, Moroi J, Wong GKC, Abrigo JM, Igase K, Matsumoto K, Wermer MJH, van Walderveen MAA, Algra A, Vergouwen MDI. ELAPSS score for prediction of risk of growth of unruptured intracranial aneurysms. Neurology.2017 Apr 25;88(17):1600-1606. doi: 10.1212 / WNL.0000000000003865. Epub 2017 Mar 31. PMID: 28363976.

[0248] Zakeri M, Atef A, Aziznia M, Jafari A. A comprehensive investigation of morphological features responsible for cerebral aneurysm rupture using machine learning. Sci Rep.2024 Jul 9;14(1):15777. doi: 10.1038 / s41598-024-66840-1. PMID: 38982160; PMCID: PMC11233616.

[0249] Tjoumakaris SI, Hanel R, Mocco J, Ali-Aziz Sultan M, Froehler M, Lieber BB, Coon A, Tateshima S, Altschul DJ, Narayanan S, El Naamani K, Taussky P, Hoh BL, Meyers P, Gounis MJ, Liebeskind DS, Volovici V, Toth G, Arthur A, Wakhloo AK; ARISE I Consortium. ARISE I Consensus Review on the Management of Intracranial Aneurysms. Stroke.2024 May;55(5):1428-1437. doi: 10.1161 / STROKEAHA.123.046208. Epub 2024 Apr 22. PMID: 38648283.

[0250] Pettersson SD, Salih M, Young M, Shutran M, Taussky P, Ogilvy CS. Predictors for Rupture of Small (<7mm) Intracranial Aneurysms: A Systematic Review and Meta- - 58 - 4224511.v11440.2081001 Analysis. World Neurosurg.2024 Feb;182:184-192.e14. doi: 10.1016 / j.wneu.2023.11.126. Epub 2023 Nov 30. PMID: 38042294.

[0251] Ou C, Qian Y, Chong W, Hou X, Zhang M, Zhang X, Si W, Duan CZ. A deep learning-based automatic system for intracranial aneurysms diagnosis on three-dimensional digital subtraction angiographic images. Med Phys.2022 Nov;49(11):7038-7053. doi: 10.1002 / mp.15846. Epub 2022 Jul 15. PMID: 35792717.

[0252] Shamshad F, Khan S, Zamir SW, Khan MH, Hayat M, Khan FS, Fu H. Transformers in medical imaging: A survey. Med Image Anal.2023 Aug;88:102802. doi: 10.1016 / j.media.2023.102802. Epub 2023 Apr 5. PMID: 37315483.

[0253] Brugnara G, Baumgartner M, Scholze ED, Deike-Hofmann K, Kades K, Scherer J, Denner S, Meredig H, Rastogi A, Mahmutoglu MA, Ulfert C, Neuberger U, Schönenberger S, Schlamp K, Bendella Z, Pinetz T, Schmeel C, Wick W, Ringleb PA, Floca R, Möhlenbruch M, Radbruch A, Bendszus M, Maier-Hein K, Vollmuth P. Deep-learning based detection of vessel occlusions on CT-angiography in patients with suspected acute ischemic stroke. Nat Commun.2023 Aug 15;14(1):4938. doi: 10.1038 / s41467-023-40564-8. PMID: 37582829; PMCID: PMC10427649.

[0254] Dhar S, Tremmel M, Mocco J, Kim M, Yamamoto J, Siddiqui AH, Hopkins LN, Meng H. Morphology parameters for intracranial aneurysm rupture risk assessment. Neurosurgery.2008 Aug;63(2):185-96; discussion 196-7. doi: 10.1227 / 01.NEU.0000316847.64140.81. PMID: 18797347; PMCID: PMC2570753.

[0255] You W, Feng J, Lu J, Chen T, Liu X, Wu Z, Gong G, Sui Y, Wang Y, Zhang Y, Ye W, Chen X, Lv J, Wei D, Tang Y, Deng D, Gui S, Lin J, Chen P, Wang Z, Gong W, Wang Y, Zhu C, Zhang Y, Saloner DA, Mitsouras D, Guan S, Li Y, Jiang Y, Wang Y. Diagnosis of intracranial aneurysms by computed tomography angiography using deep learning-based detection and segmentation. J Neurointerv Surg.2024 Dec 26;17(e1):e132- e138. doi: 10.1136 / jnis-2023-021022. PMID: 38238009.

[0256] U.S. Food and Drug Administration.510(k) premarket notification: BriefCase. [Internet]. Updated 2024 Oct 7 [cited 2024 Oct 11]. Available from: www.accessdata.fda.gov / scripts / cdrh / cfdocs / cfpmn / pmn.cfm?ID=K213721.

[0257] U.S. Food and Drug Administration.510(k) premarket notification: Viz ANEURYSM, Viz ANX. [Internet]. Updated 2024 Oct 7 [cited 2024 Oct 11]. Available from: www.accessdata.fda.gov / scripts / cdrh / cfdocs / cfpmn / pmn.cfm?ID=K213319. - 59 - 4224511.v11440.2081001

[0258] U.S. Food and Drug Administration.510(k) premarket notification: Rapid Aneurysm Triage and Notification. [Internet]. Updated 2024 Oct 7 [cited 2024 Oct 11]. Available from: www.accessdata.fda.gov / scripts / cdrh / cfdocs / cfpmn / pmn.cfm?ID=K230074.

[0259] Kim H, Ballekere A, Ali I, Marioni S, Abdelkhaleq R, Niktabe A, Azeem H, Iyyangar A, Segev O, Bibas O, Paz D, Love C, Giancardo L, Sheth S. Machine Learning– Enabled Detection of Unruptured Cerebral Aneurysms Improves Detection Rates and Clinical Care. Stroke: Vascular and Interventional Neurology.2023 Sep 13;3(6). doi:10.1161 / svin.123.000938.

[0260] UCAS Japan Investigators; Morita A, Kirino T, Hashi K, Aoki N, Fukuhara S, Hashimoto N, Nakayama T, Sakai M, Teramoto A, Tominari S, Yoshimoto T. The natural course of unruptured cerebral aneurysms in a Japanese cohort. N Engl J Med.2012 Jun 28;366(26):2474-82. doi: 10.1056 / NEJMoa1113260. PMID: 22738097.

[0261] Nam HH, Jang DK, Cho BR. Complications and risk factors after digital subtraction angiography: 1-year single-center study. J Cerebrovasc Endovasc Neurosurg. 2022 Dec;24(4):335-340. doi: 10.7461 / jcen.2022.E2022.05.001. Epub 2022 Sep 26. PMID: 36153862; PMCID: PMC9829562.

[0262] Yang K, Musio F, Ma Y, Juchler N, Paetzold JC, Al-Maskari R, Höher L, Li HB, Hamamci IE, Sekuboyina A, Shit S, Huang H, Prabhakar C, de la Rosa E, Waldmannstetter D, Kofler F, Navarro F, Menten M, Ezhov I, Rueckert D, Vos I, Ruigrok Y, Velthuis B, Kuijf H, Hämmerli J, Wurster C, Bijlenga P, Westphal L, Bisschop J, Colombo E, Baazaoui H, Makmur A, Hallinan J, Wiestler B, Kirschke JS, Wiest R, Montagnon E, Letourneau-Guillon L, Galdran A, Galati F, Falcetta D, Zuluaga MA, Lin C, Zhao H, Zhang Z, Ra S, Hwang J, Park H, Chen J, Wodzinski M, Müller H, Shi P, Liu W, Ma T, Yalçin C, Hamadache RE, Salvi J, Llado X, Lal-Trehan Estrada UM, Abramova V, Giancardo L, Oliver A, Liu J, Huang H, Cui Y, Lin Z, Liu Y, Zhu S, Patel TR, Tutino VM, Orouskhani M, Wang H, Mossa-Basha M, Zhu C, Rokuss MR, Kirchhoff Y, Disch N, Holzschuh J, Isensee F, Maier-Hein K, Sato Y, Hirsch S, Wegener S, Menze B. Benchmarking the CoW with the TopCoW Challenge: Topology-Aware Anatomical Segmentation of the Circle of Willis for CTA and MRA. ArXiv [Preprint].2024 Apr 29:arXiv:2312.17670v3. PMID: 38235066; PMCID: PMC10793481. - 60 - 4224511.v1

Claims

1440.2081001 CLAIMS What is claimed is:

1. A computer-implemented method of autonomous deep learning morphology analysis, the computer-implemented method comprising: obtaining at least one image associated with a patient; processing the obtained at least one image using an autonomous deep learning model to extract at least one morphological parameter associated with a suspected pathology of the patient; and based on the extracted at least one morphological parameter, generating an output analysis; said obtaining, processing, and generating being automatically performed by one or more digital processors.

2. The computer-implemented method of Claim 1, wherein the suspected pathology is one or more aneurysms, and wherein generating the output analysis by the one or more digital processors includes at least one of: (i) automatically detecting existence and location of the one or more aneurysms, (ii) responsively generating a three-dimensional (3D) reconstruction of vasculature corresponding to the detected one or more aneurysms, and (iii) automatically generating a morphology characterization of the detected one or more aneurysms.

3. The computer-implemented method of Claim 2, wherein the generated morphology characterization includes, for each detected aneurysm, indications of any combination of: volume, surface area, neck width, dome width, dome height, dome depth, proximal parent artery diameter, distal parent artery diameter, flow angle, undulation index, ellipticity index, non-sphericity index, aspect ratio, size ratio, width-parent artery diameter ratio, and height-neck ratio.

4. The computer-implemented method of Claim 2, wherein the vasculature is intracranial vasculature.

5. The computer-implemented method of Claim 2, further comprising: - 61 - 4224511.v11440.2081001 automatically determining, by the one or more digital processors, a rupture probability of the detected one or more aneurysms.

6. The computer-implemented method of Claim 2, further comprising: transforming, by the one or more digital processors, the generated 3D reconstruction of vasculature into a volume mesh; automatically determining, by the one or more digital processors: (i) at least one boundary condition for the volume mesh, and (ii) at least one fluid property for the volume mesh; responsively performing, by the one or more digital processors, a fluid dynamics simulation based on the volume mesh, the determined at least one boundary condition, and the determined at least one fluid property; and automatically extracting, by the one or more digital processors, at least one hemodynamic feature based on a result of performing the fluid dynamics simulation.

7. The computer-implemented method of Claim 6, wherein transforming the generated 3D reconstruction of vasculature includes, by the one or more digital processors: generating a surface mesh based on the generated 3D reconstruction of vasculature; and transforming the generated surface mesh into the volume mesh.

8. The computer-implemented method of Claim 7, wherein generating the surface mesh is based on a surface smoothing technique.

9. The computer-implemented method of Claim 8, wherein the surface smoothing technique is at least one of Laplacian surface smoothing and Taubin surface smoothing.

10. The computer-implemented method of Claim 6, further comprising: based on one or more quality metrics of the volume mesh, automatically determining at least one of: (i) numerical stability of the volume mesh, and (ii) numerical accuracy of the volume mesh.

11. The computer-implemented method of Claim 10, wherein the one or more quality metrics include at least one of: aspect ratio, skewness, and minimum angle. - 62 - 4224511.v11440.2081001 12. The computer-implemented method of Claim 6, wherein the volume mesh includes at least one tetrahedral element that defines an interior fluid domain of the vasculature.

13. The computer-implemented method of Claim 6, wherein the determined at least one boundary condition includes at least one of: (i) one or more inlet conditions, and (ii) one or more outlet conditions.

14. The computer-implemented method of Claim 13, further comprising: automatically determining the one or more inlet conditions based on at least one of: (i) one or more patient-specific velocity profiles, and (ii) one or more standardized pulsatile waveforms.

15. The computer-implemented method of Claim 13, further comprising: automatically determining the one or more outlet conditions based on at least one of: (i) one or more fixed pressure models, and (ii) one or more resistance-based parameter models.

16. The computer-implemented method of Claim 15, wherein the one or more resistance- based parameter models include a Windkessel model.

17. The computer-implemented method of Claim 6, wherein automatically determining the at least one fluid property is based on a Newtonian blood model or a non- Newtonian blood model.

18. The computer-implemented method of Claim 6, wherein the determined at least one fluid property includes at least one of: density and dynamic viscosity.

19. The computer-implemented method of Claim 6, wherein the fluid dynamics simulation is performed using a time-resolved model or a steady-state model.

20. The computer-implemented method of Claim 6, wherein the extracted at least one hemodynamic feature includes at least one of: (i) wall shear stress (WSS), (ii) an oscillatory shear index (OSI), (iii) time-averaged WSS (TAWSS), (iv) a flow jet velocity, (v) a flow jet directionality, and (vi) intra-aneurysmal pressure.

21. The computer-implemented method of Claim 6, further comprising: - 63 - 4224511.v11440.2081001 automatically correlating, by the one or more digital processors, the extracted at least one hemodynamic feature with the obtained at least one image.

22. The computer-implemented method of Claim 6, further comprising: automatically determining, by the one or more digital processors, vessel wall elasticity for the volume mesh; wherein performing the fluid dynamics simulation includes performing a fluid-structure interaction (FSI) simulation based on the determined vessel wall elasticity.

23. The computer-implemented method of Claim 6, wherein: the fluid dynamics simulation is performed by the one or more digital processors using a deep learning-based surrogate model; and the at least one hemodynamic feature is extracted in real-time of the fluid dynamics simulation.

24. The computer-implemented method of Claim 6, further comprising: automatically determining, by the one or more digital processors, a rupture probability of the detected one or more aneurysms based on the extracted at least one hemodynamic feature.

25. The computer-implemented method of Claim 1, wherein the autonomous deep learning model is a Brain Aneurysm Morphology (BAM) model.

26. The computer-implemented method of Claim 1, wherein the obtained at least one image includes at least one of: a computed tomography (CT) scan and a magnetic resonance imaging (MRI) scan.

27. A computer-based system for autonomous deep learning morphology analysis, the computer-based system comprising: at least one processor; and a memory with computer code instructions stored thereon, the at least one processor and the memory, with the computer code instructions, being configured to cause the computer-based system to: obtain at least one image associated with a patient; - 64 - 4224511.v11440.2081001 process the obtained at least one image using an autonomous deep learning model to extract at least one morphological parameter associated with a suspected pathology of the patient; and based on the at least one morphological parameter, generate an output analysis.

28. The computer-based system of Claim 27, wherein the suspected pathology is one or more aneurysms, and where, in generating the output analysis, the at least one processor and the memory, with the computer code instructions, are configured to cause the computer-based system to perform at least one of: (i) automatically detecting existence and location of the one or more aneurysms, (ii) responsively generating a three-dimensional (3D) reconstruction of vasculature corresponding to the detected one or more aneurysms, and (iii) automatically generating a morphology characterization of the detected one or more aneurysms.

29. The computer-based system of Claim 28, wherein the generated morphology characterization includes, for each detected aneurysm, indications of any combination of: volume, surface area, neck width, dome width, dome height, dome depth, proximal parent artery diameter, distal parent artery diameter, flow angle, undulation index, ellipticity index, non-sphericity index, aspect ratio, size ratio, width-parent artery diameter ratio, and height-neck ratio.

30. The computer-based system of Claim 28, wherein the vasculature is intracranial vasculature 31. The computer-based system of Claim 28, wherein the at least one processor and the memory, with the computer code instructions, are further configured to cause the computer-based system to: determine a rupture probability of the detected one or more aneurysms.

32. The computer-based system of Claim 28, wherein the at least one processor and the memory, with the computer code instructions, are further configured to cause the computer-based system to: - 65 - 4224511.v11440.2081001 transform the generated 3D reconstruction of vasculature into a volume mesh; automatically determine: (i) at least one boundary condition for the volume mesh, and (ii) at least one fluid property for the volume mesh; responsively perform a fluid dynamics simulation based on the volume mesh, the determined at least one boundary condition, and the determined at least one fluid property; and automatically extract at least one hemodynamic feature based on a result of the performed fluid dynamics simulation.

33. The computer-based system of Claim 32, where, in transforming the generated 3D reconstruction of vasculature, the at least one processor and the memory, with the computer code instructions, are configured to cause the computer-based system to: generate a surface mesh based on the generated 3D reconstruction of vasculature; and transforming the generated surface mesh into the volume mesh.

34. The computer-based system of Claim 33, wherein the surface mesh is generated based on a surface smoothing technique.

35. The computer-based system of Claim 34, wherein the surface smoothing technique is at least one of Laplacian surface smoothing and Taubin surface smoothing.

36. The computer-based system of Claim 32, wherein the at least one processor and the memory, with the computer code instructions, are further configured to cause the computer-based system to: based on one or more quality metrics of the volume mesh, automatically determine at least one of: (i) numerical stability of the volume mesh, and (ii) numerical accuracy of the volume mesh.

37. The computer-based system of Claim 36, wherein the one or more quality metrics include at least one of: aspect ratio, skewness, and minimum angle.

38. The computer-based system of Claim 32, wherein the volume mesh includes at least one tetrahedral element that defines an interior fluid domain of the vasculature. - 66 - 4224511.v11440.2081001 39. The computer-based system of Claim 32, wherein the determined at least one boundary condition includes at least one of: (i) one or more inlet conditions, and (ii) one or more outlet conditions.

40. The computer-based system of Claim 39, wherein the at least one processor and the memory, with the computer code instructions, are further configured to cause the computer-based system to: automatically determine the one or more inlet conditions based on at least one of: (i) one or more patient-specific velocity profiles, and (ii) one or more standardized pulsatile waveforms.

41. The computer-based system of Claim 39, wherein the at least one processor and the memory, with the computer code instructions, are further configured to cause the computer-based system to: automatically determine the one or more outlet conditions based on at least one of: (i) one or more fixed pressure models, and (ii) one or more resistance-based parameter models.

42. The computer-based system of Claim 41, wherein the one or more resistance-based parameter models include a Windkessel model.

43. The computer-based system of Claim 32, wherein the at least one fluid property is automatically determined based on a Newtonian blood model or a non-Newtonian blood model.

44. The computer-based system of Claim 32, wherein the determined at least one fluid property includes at least one of: density and dynamic viscosity.

45. The computer-based system of Claim 32, wherein the fluid dynamics simulation is performed using a time-resolved model or a steady-state model.

46. The computer-based system of Claim 32, wherein the extracted at least one hemodynamic feature includes at least one of: (i) wall shear stress (WSS), (ii) an oscillatory shear index (OSI), (iii) time-averaged WSS (TAWSS), (iv) a flow jet velocity, (v) a flow jet directionality, and (vi) intra-aneurysmal pressure. - 67 - 4224511.v11440.2081001 47. The computer-based system of Claim 32, wherein the at least one processor and the memory, with the computer code instructions, are further configured to cause the computer-based system to: automatically correlate the extracted at least one hemodynamic feature with the obtained at least one image.

48. The computer-based system of Claim 32, wherein the at least one processor and the memory, with the computer code instructions, are further configured to cause the computer-based system to: automatically determine vessel wall elasticity for the volume mesh; and where, in performing the fluid dynamics simulation, the at least one processor and the memory, with the computer code instructions, are configured to cause the computer-based system to perform a fluid-structure interaction (FSI) simulation based on the determined vessel wall elasticity.

49. The computer-based system of Claim 32, wherein: the fluid dynamics simulation is performed using a deep learning-based surrogate model; and the at least one hemodynamic feature is extracted in real-time of the fluid dynamics simulation.

50. The computer-based system of Claim 32, wherein the at least one processor and the memory, with the computer code instructions, are further configured to cause the computer-based system to: automatically determine a rupture probability of the detected one or more aneurysms based on the extracted at least one hemodynamic feature.

51. The computer-based system of Claim 27, wherein the autonomous deep learning model is a Brain Aneurysm Morphology (BAM) model 52. The computer-based system of Claim 27, wherein the obtained at least one image includes at least one of: a computed tomography (CT) scan and a magnetic resonance imaging (MRI) scan. - 68 - 4224511.v11440.2081001 53. A computer program product for autonomous deep learning morphology analysis, the computer program product comprising a non-transitory computer-readable medium with computer code instructions stored thereon, the computer code instructions being configured, when executed by a processor, to cause an apparatus associated with the processor to: obtain at least one image associated with a patient; process the obtained at least one image using an autonomous deep learning model to extract at least one morphological parameter associated with a suspected pathology of the patient; and based on the at least one morphological parameter, generate an output analysis.

54. The computer program product of Claim 53, wherein the suspected pathology is one or more aneurysms, and where, in generating the output analysis, the computer code instructions, when executed by the processor, cause the apparatus associated with the processor to perform at least one of: (i) automatically detecting existence and location of the one or more aneurysms, (ii) responsively generating a three-dimensional (3D) reconstruction of vasculature corresponding to the detected one or more aneurysms, and (iii) automatically generating a morphology characterization of the detected one or more aneurysms.

55. The computer program product of Claim 54, wherein the generated morphology characterization includes, for each detected aneurysm, indications of any combination of: volume, surface area, neck width, dome width, dome height, dome depth, proximal parent artery diameter, distal parent artery diameter, flow angle, undulation index, ellipticity index, non-sphericity index, aspect ratio, size ratio, width-parent artery diameter ratio, and height-neck ratio.

56. The computer program product of Claim 54, wherein the vasculature is intracranial vasculature. - 69 - 4224511.v11440.2081001 57. The computer program product of Claim 54, wherein the computer code instructions are further configured, when executed by the processor, to cause the apparatus associated with the processor to: determine a rupture probability of the detected one or more aneurysms.

58. The computer program product of Claim 54, wherein the computer code instructions are further configured, when executed by the processor, to cause the apparatus associated with the processor to: transform the generated 3D reconstruction of vasculature into a volume mesh; automatically determine: (i) at least one boundary condition for the volume mesh, and (ii) at least one fluid property for the volume mesh; responsively perform a fluid dynamics simulation based on the volume mesh, the determined at least one boundary condition, and the determined at least one fluid property; and automatically extract at least one hemodynamic feature based on a result of the performed fluid dynamics simulation.

59. The computer program product of Claim 58, where, in transforming the generated 3D reconstruction of vasculature, the computer code instructions, when executed by the processor, cause the apparatus associated with the processor to: generate a surface mesh based on the generated 3D reconstruction of vasculature; and transforming the generated surface mesh into the volume mesh.

60. The computer program product of Claim 59, wherein the surface mesh is generated based on a surface smoothing technique.

61. The computer program product of Claim 60, wherein the surface smoothing technique is at least one of Laplacian surface smoothing and Taubin surface smoothing.

62. The computer program product of Claim 58, wherein the computer code instructions are further configured, when executed by the processor, to cause the apparatus associated with the processor to: - 70 - 4224511.v11440.2081001 based on one or more quality metrics of the volume mesh, automatically determine at least one of: (i) numerical stability of the volume mesh, and (ii) numerical accuracy of the volume mesh.

63. The computer program product of Claim 62, wherein the one or more quality metrics include at least one of: aspect ratio, skewness, and minimum angle.

64. The computer program product of Claim 58, wherein the volume mesh includes at least one tetrahedral element that defines an interior fluid domain of the vasculature.

65. The computer program product of Claim 58, wherein the determined at least one boundary condition includes at least one of: (i) one or more inlet conditions, and (ii) one or more outlet conditions.

66. The computer program product of Claim 65, wherein the computer code instructions are further configured, when executed by the processor, to cause the apparatus associated with the processor to: automatically determine the one or more inlet conditions based on at least one of: (i) one or more patient-specific velocity profiles, and (ii) one or more standardized pulsatile waveforms.

67. The computer program product of Claim 65, wherein the computer code instructions are further configured, when executed by the processor, to cause the apparatus associated with the processor to: automatically determine the one or more outlet conditions based on at least one of: (i) one or more fixed pressure models, and (ii) one or more resistance-based parameter models.

68. The computer program product of Claim 67, wherein the one or more resistance- based parameter models include a Windkessel model.

69. The computer program product of Claim 58, wherein the at least one fluid property is automatically determined based on a Newtonian blood model or a non-Newtonian blood model. - 71 - 4224511.v11440.2081001 70. The computer program product of Claim 58, wherein the determined at least one fluid property includes at least one of: density and dynamic viscosity.

71. The computer program product of Claim 58, wherein the fluid dynamics simulation is performed using a time-resolved model or a steady-state model.

72. The computer program product of Claim 58, wherein the extracted at least one hemodynamic feature includes at least one of: (i) wall shear stress (WSS), (ii) an oscillatory shear index (OSI), (iii) time-averaged WSS (TAWSS), (iv) a flow jet velocity, (v) a flow jet directionality, and (vi) intra-aneurysmal pressure.

73. The computer program product of Claim 58, wherein the computer code instructions are further configured, when executed by the processor, to cause the apparatus associated with the processor to: automatically correlate the extracted at least one hemodynamic feature with the obtained at least one image.

74. The computer program product of Claim 58, wherein the computer code instructions are further configured, when executed by the processor, to cause the apparatus associated with the processor to: automatically determine vessel wall elasticity for the volume mesh; and where, in performing the fluid dynamics simulation, the at least one processor and the memory, with the computer code instructions, are configured to cause the computer-based system to perform a fluid-structure interaction (FSI) simulation based on the determined vessel wall elasticity.

75. The computer program product of Claim 58, wherein: the fluid dynamics simulation is performed using a deep learning-based surrogate model; and the at least one hemodynamic feature is extracted in real-time of the fluid dynamics simulation.

76. The computer program product of Claim 58, wherein the computer code instructions are further configured, when executed by the processor, to cause the apparatus associated with the processor to: - 72 - 4224511.v11440.2081001 automatically determine a rupture probability of the detected one or more aneurysms based on the extracted at least one hemodynamic feature.

77. The computer program product of Claim 53, wherein the autonomous deep learning model is a Brain Aneurysm Morphology (BAM) model.

78. The computer program product of Claim 53, wherein the obtained at least one image includes at least one of: a computed tomography (CT) scan and a magnetic resonance imaging (MRI) scan. - 73 - 4224511.v1

Images