There is a massive amount of clinical three-dimensional (3D) cardiac image data available today in numerous hospitals, but such data has been considerably underutilized in both clinical and engineering analyses of cardiac function. These 3D data offers unique and valuable information, allowing researchers to develop innovative, personalized approaches to treat diseases. Furthermore, using these 3D datasets as input to computational models can facilitate a population-based analysis that can be used to quantify uncertainty in treatment procedures, and can be utilized for virtual clinical trials for innovative device development. However, there are several critical technical bottlenecks preventing simulation-based clinical evaluation a reality: 1) difficulty in automatic 3D reconstruction of thin complex structures such as heart valve leaflets from clinical images, 2) computational models are constructed without mesh correspondence, which makes it challenging to run batch simulations and conduct large patient population data analyses due to inconsistencies in model setups, and 3) computing time is long, which inhibits prompt feedback for clinical use. A potential paradigm-changing solution to the challenges is to incorporate machine learning algorithms to expedite the geometry reconstruction and computational analysis procedures. Therefore, the objective of this proposal is to develop a novel computing framework, using advanced tissue modeling and machine learning techniques, to automatically process pre-operative clinical image data and predict post-operative clinical outcomes. Transcatheter aortic valve replacement (TAVR) intervention will serve as a testbed for the modeling methods.
In Aim 1, we will develop novel shape dictionary learning (SDL) based methods for automatic reconstruction of TAVR patient aortic valves. Through the modeling process, mesh correspondence will be established across the patient geometric models. The distribution and variation of TAVR patient geometries will be described by statistical shape models (SSMs).
In Aim 2, population-based FE analysis of the TAVR procedure will be conducted on thousands of virtual patient models generated by the SSMs (Aim 1). A deep neural network (DNN) will be developed and trained to learn the relationship between the TAVR FE inputs and outputs. Successful completion of this study will result in a ML-FE surrogate for TAVR analysis, combining the automated TAVR patient geometry reconstruction algorithms and the trained DNN, to provide fast TAVR biomechanics analysis without extensive re-computing of the model. Furthermore, the algorithms developed in this study can be generalized for other applications and devices.
Current clinical image modalities can be utilized to develop patient-specific computational models to pre-operatively plan transcatheter aortic valve replacement (TAVR) procedures. However, the computational modeling and simulation processes are time-consuming, which limits clinical translatability. Thus, the objective of this proposal is to develop algorithms using machine learning techniques to rapidly process and predict TAVR computational simulation outcomes directly from clinical image data.
