Computation of a probabilistic and anisotropic failure metric on the aortic wall using a machine learning-based surrogate model

https://doi.org/10.1016/j.compbiomed.2021.104794Get rights and content

Highlights

  • End-to-end failure metric computation enabled by a machine learning model.

  • Effect of hyperelastic properties is incorporated using two-phase aorta geometries.

  • CT images and tissue testing data of real patients are used to generate FEA data.

  • The FP failure metric can be computed with 0.42% NMAE in 1 s.

  • The FP have more discriminative power than deterministic or isotropic metrics.

Abstract

Scalar-valued failure metrics are commonly used to assess the risk of aortic aneurysm rupture and dissection, which occurs under hypertensive blood pressures brought on by extreme emotional or physical stress. To compute failure metrics under an elevated blood pressure, a classical patient-specific computer model consists of multiple computation steps involving inverse and forward analyses. These classical procedures may be impractical for time-sensitive clinical applications that require prompt feedback to clinicians. In this study, we developed a machine learning-based surrogate model to directly predict a probabilistic and anisotropic failure metric, namely failure probability (FP), on the aortic wall using aorta geometries at the systolic and diastolic phases. Ascending thoracic aortic aneurysm (ATAA) geometries of 60 patients were obtained from their CT scans, and biaxial mechanical testing data of ATAA tissues from 79 patients were collected. Finite element simulations were used to generate datasets for training, validation, and testing of the ML-surrogate model. The testing results demonstrated that the ML-surrogate can compute the maximum FP failure metric, with 0.42% normalized mean absolute error, in 1 s. To compare the performance of the ML-predicted probabilistic FP metric with other isotropic or deterministic metrics, a numerical case study was performed using synthetic “baseline” data. Our results showed that the probabilistic FP metric had more discriminative power than the deterministic Tsai-Hill metric, isotropic maximum principal stress, and aortic diameter criterion.

Introduction

Thoracic aortic aneurysm (TAA) is a life-threatening condition that can lead to rupture and dissection of the aorta. If left untreated, the five-year survival rate is 54% [1]. The pathology of the thoracic aorta can be eradicated by elective surgery or endovascular treatment. Aortic size-based criterion is currently used to triage patients to surgery, which is often recommended for patients with TAA diameter larger than 5–5.5 cm [2,3]. However, the aortic size may not accurately reflect the risk for all patients [4,5]: some small aneurysms (e.g., <4 cm) have been known to rupture [5]. As adverse events of aneurysms (i.e., rupture and dissection) are mechanics-driven, aortic wall stresses, computed from patient-specific finite element analysis (FEA) models, can provide more valuable insights into TAA risk [6,7]. To assess the risk of TAA rupture and dissection, the FEA-computed stress components are typically converted into a scalar-valued index (i.e., failure metric).

Isotropic failure indices are often employed in patient-specific models to assess the risk of aortic aneurysms, including the von Mises equivalent stress [8], the maximum principal stress (MPS) [9], and rupture potential index (RPI) [10]. Since failure strengths of aortic tissue are highly directional dependent (i.e., anisotropic) [[11], [12], [13], [14], [15], [16]], these isotropic metrics may not be appropriate for the aortic wall [17]. Experimental works have also revealed large variations of aortic wall strengths [11,13,14,16], which suggest that the probability distribution of wall strengths can be incorporated for a more accurate failure metric. In a recent study [18], we proposed a probabilistic and anisotropic failure metric, namely the failure probability (FP), by using tissue failure testing data of 84 ascending thoracic aortic aneurysm (ATAA) patients. Using synthetic data, the novel FP metric demonstrated a better discriminative capability than that of isotropic or deterministic failure metrics.

Since TAA dissection/rupture usually occurs under hypertensive blood pressures brought on by extreme emotional or physical stress [19,20], failure metric is typically evaluated under an elevated blood pressure in patient-specific FEA models [7,21]. In a classical patient-specific model, computation of failure metrics consists of the following steps: (1) segmentation, registration, and mesh generation of the aortic wall at multiple phases (e.g., systole and diastole) from clinical images; (2) identification of unknown patient-specific hyperelastic material parameters from multi-phase aorta geometries by using an optimization-based inverse method [[22], [23], [24], [25], [26]]; (3) computing wall stresses at an elevated blood pressure using nonlinear FEA simulation [7,21]; (4) post-processing to convert the FEA-computed stress distributions into a scalar-valued failure metric distribution; and (5) capturing the maximum failure metric on the aortic wall. However, these classical procedures may not be suitable for clinical applications that require prompt feedback because: (i) manual segmentation and mesh generation requires heavy human involvements, which are labor-intensive and time-consuming; (ii) optimization-based inverse methods are iterative in nature and thus computationally expensive, which can take hours to weeks to complete for a single patient [[22], [23], [24], [25], [26]]; (iii) FEA stress computation and post-processing can also take minutes to hours for one patient; and (iv) convergence issues can occur in nonlinear FEA simulations. Hence, the computation of failure metrics using the classical methods is impractical in time-sensitive clinical settings.

Machine learning (ML) and deep learning (DL) approaches have led to great success in many real-world applications [[27], [28], [29], [30], [31]]. Recently, DL-based algorithms have shed light on automatic image segmentation and mesh generation in real-time [29,[32], [33], [34]]. To alleviate the high computational cost of classical methods, data-driven techniques, in particular ML algorithms, have been recently employed as surrogate models to accelerate traditional computation procedures. For instance, deep learning approaches were proposed by our group and others as fast and accurate surrogates of FEA-based stress analysis [[35], [36], [37]] and computational fluid dynamics (CFD) based hemodynamic analysis [38]. ML-models were developed to predict abdominal aortic aneurysm expansion as surrogates of growth and remodeling approaches [39,40]. In a previous study [41], we developed an ML approach to rapidly (i.e., within 1 s) identify hyperelastic parameters from two-phase aorta geometries. An ML-based surrogate model can be trained to establish the nonlinear and complex relationship between inputs and outputs, which can fundamentally solve the computational cost challenge. By supervised training on a large FEA simulation dataset, an ML-surrogate model may be developed to compute failure metrics directly from multi-phase aorta geometries, bypassing the abovementioned classical steps (2)–(5). Once trained, the ML-surrogate can enable real-time prediction of the failure metric.

In this paper, we developed an ML-based surrogate model to compute the probabilistic and anisotropic FP metric. As shown in Fig. 1, the ML-model used the same inputs as in the classical patient-specific FEA model, i.e., the aorta geometries at the systolic and diastolic phases. Our previously proposed ML framework [41] was extended for direct failure metric computation, and the hyperelastic material parameter identification, as well as the nonlinear FEA simulation, can be bypassed. The systolic and diastolic geometries were encoded into shape codes using principal component analysis. The failure metrics at different quantiles were predicted from shape codes using a fully connected neural network. CT-derived geometries of 60 ATAA patients and tissue testing data of 79 patients were obtained. FEA simulations were used to generate datasets for training, validation, and testing of the ML-surrogate model. The predictive accuracy of the ML-surrogate was examined. To compare the performance of the ML-predicted FP metrics with other failure metrics, a numerical case study was performed using synthetic “baseline” data.

Section snippets

Representative ATAA geometries and material properties

To train, validate, and test the ML-surrogate model, FEA simulation data were needed. By following the same procedures established in Ref. [41], we first generated representative aorta geometries from clinical images of real patients by using a statistical shape model (SSM). We sampled representative sets of hyperelastic parameters from experimental tissue testing data by constructing a convex hull in the material parameter space. We extended the methods in Ref. [41] by incorporating CT-derived

Prediction of the FP metric

The number of units of the neural network was determined using grid search. Ten-fold cross-validations were performed to evaluate the NMAE and NSTAE with 3000 epochs, the results are shown in Table 2. The lowest NMAE was achieved by using 2 hidden layers and 256 units in each layer.

Using the 15623 training and validation cases, the neural network was trained with 10,000 epochs in about 30 min on a single GPU. Given a pair of systolic and diastolic geometries as the inputs, the trained ML-model

Discussion

In this study, we developed a novel end-to-end ML approach that enabled fast and accurate prediction of a probabilistic failure metric on the aortic wall. Comparing with classical procedures to obtain the FP metric (i.e., optimization-based material parameter identification [[22], [23], [24], [25], [26]] and nonlinear FEA [7,21]), the ML approach can fundamentally resolve the computation cost challenge. Using aorta geometries at two cardiac phases, different quantiles of the FP metric under

Conclusion

We proposed an end-to-end ML-based surrogate model to predict a probabilistic and anisotropic failure (FP) metric on the aortic wall from in vivo geometries at two cardiac phases and given blood pressures. FEA simulations were employed to generate the training, validation, and testing datasets. The proposed ML-based predictive model accelerated the computation of failure metrics, bypassing the traditional process that requires both inverse identification and forward computation. The trained

Declaration of competing interest

Dr. Wei Sun serves as the Chief Scientific Advisor of Dura Biotech. He has received compensation and owns equity in the company.

Acknowledgements

This study is supported by American Heart Association (AHA) 18TPA34230083.

References (73)

  • M. Liu et al.

    A probabilistic and anisotropic failure metric for ascending thoracic aortic aneurysm risk assessment

    J. Mech. Phys. Solid.

    (2021)
  • I.S. Hatzaras et al.

    Role of exertion or emotion as inciting events for acute aortic dissection

    Am. J. Cardiol.

    (2007)
  • M. Liu et al.

    A new inverse method for estimation of in vivo mechanical properties of the aortic wall

    Journal of the mechanical behavior of biomedical materials

    (2017)
  • M. Liu et al.

    Estimation of in vivo mechanical properties of the aortic wall: a multi-resolution direct search approach

    Journal of the mechanical behavior of biomedical materials

    (2018)
  • A. Wittek et al.

    A finite element updating approach for identification of the anisotropic hyperelastic properties of normal and diseased aortic walls from 4D ultrasound strain imaging

    Journal of the Mechanical Behavior of Biomedical Materials

    (2016)
  • A. Wittek et al.

    In vivo determination of elastic properties of the human aorta based on 4D ultrasound data

    Journal of the mechanical behavior of biomedical materials

    (2013)
  • L. Liang et al.

    A feasibility study of deep learning for predicting hemodynamics of human thoracic aorta

    J. Biomech.

    (2020)
  • Z. Jiang et al.

    Machine learning approaches to surrogate multifidelity Growth and Remodeling models for efficient abdominal aortic aneurysmal applications

    Comput. Biol. Med.

    (2021)
  • M. Liu et al.

    Estimation of in vivo constitutive parameters of the aortic wall using a machine learning approach

    Comput. Methods Appl. Mech. Eng.

    (2019)
  • M. Leshno et al.

    Multilayer feedforward networks with a nonpolynomial activation function can approximate any function

    Neural Network.

    (1993)
  • K. Hornik

    Approximation capabilities of multilayer feedforward networks

    Neural Network.

    (1991)
  • X. He et al.

    Prediction of local strength of ascending thoracic aortic aneurysms

    Journal of the Mechanical Behavior of Biomedical Materials

    (2021)
  • M. Raissi et al.

    Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    J. Comput. Phys.

    (2019)
  • E. Haghighat et al.

    A physics-informed deep learning framework for inversion and surrogate modeling in solid mechanics

    Comput. Methods Appl. Mech. Eng.

    (2021)
  • D.P. Sokolis

    Effects of aneurysm on the directional, regional, and layer distribution of residual strains in ascending thoracic aorta

    Journal of the Mechanical Behavior of Biomedical Materials

    (2015)
  • C.-Y. Liu et al.

    Evolution of aortic wall thickness: long-term follow up from the Multi-Ethnic Study of Atherosclerosis (MESA)

    J. Cardiovasc. Magn. Reson.

    (2013)
  • E.K. Shang et al.

    Validation of semiautomated and locally resolved aortic wall thickness measurements from computed tomography

    J. Vasc. Surg.

    (2015)
  • D.M. Pierce et al.

    A method for incorporating three-dimensional residual stretches/stresses into patient-specific finite element simulations of arteries

    Journal of the mechanical behavior of biomedical materials

    (2015)
  • H. Liu et al.

    Finite element simulation of three dimensional residual stress in the aortic wall using an anisotropic tissue growth model

    Journal of the mechanical behavior of biomedical materials

    (2019)
  • S. Polzer et al.

    Biomechanical indices are more sensitive than diameter in predicting rupture of asymptomatic abdominal aortic aneurysms

    J. Vasc. Surg.

    (2020)
  • G.R. Joldes et al.

    BioPARR: a software system for estimating the rupture potential index for abdominal aortic aneurysms

    Sci. Rep.

    (2017)
  • J.P. Vande Geest et al.

    A biomechanics‐based rupture potential index for abdominal aortic aneurysm risk assessment: demonstrative application

    Ann. N. Y. Acad. Sci.

    (2006)
  • Z. Teng et al.

    Layer-and direction-specific material properties, extreme extensibility and ultimate material strength of human abdominal aorta and aneurysm: a uniaxial extension study

    Ann. Biomed. Eng.

    (2015)
  • J.-H. Kim et al.

    Experimental characterization of rupture in human aortic aneurysms using a full-field measurement technique

    Biomech. Model. Mechanobiol.

    (2012)
  • M. Liu et al.

    A novel anisotropic failure criterion with dispersed fiber orientations for aortic tissues

    J. Biomech. Eng.

    (2020)
  • C.E. Korenczuk et al.

    Isotropic failure criteria are not appropriate for anisotropic fibrous biological tissues

    J. Biomech. Eng.

    (2017)
  • Cited by (0)

    View full text