Postdoctoral Research Fellow
Mohammadmostafa Asheghan, Hoda Javadikasgari, Taraneh Attary, Farhad R. Nezami
Dr. Farhad R. Nezami
Research Category: Cardiovascular, Diabetes, and Metabolic Disorders
Left ventricular mass index (LVMI) regression is an expected phenomena after transcatheter aortic valve replacement (TAVR) surgery. It has been shown that it is tightly correlated with hospitalization likelihood one year after the operation and can be a strong survival score indicator. Therefore, an estimation of LVMI regression before the surgery will be of high clinical importance. In this project, we aim to predict LVMI regression based on the pre-operation CT images of the LV.
The process starts with automatically segmentation of CT images to get the LV point clouds. After denoising and down-sampling, the shapes will be aligned through iterative closest point (ICP) algorithm. Through AI-powered techniques, order reduction on the vector is performed such that the extracted principal components out of the shapes serve the maximum correlation with the outcome vector. Using a machine learning novel approach we not only predict the LVMI regression at a specific time, but also generate patient-specific curves that provide an estimate of LVMI evolution over the time. The high accuracy of the results demonstrate that shapes convey a load of hidden information that are not visible or commonly attended in traditional clinical approaches such as measuring ventricular volume and thickness.
Narrowing of the valve in the aorta is called Aortic stenosis (AS). Complementing the surgery for old patients nowadays is a minimally invasive operation of transcatheter aortic valve replacement (TAVR). One of the adverse outcomes of TAVR is the change in heart muscle of left heart measured by left ventricular mass index (LVMI), which is called LVMI regression after the surgery. Having an estimation of LVMI change before the surgery will help the medical team to estimate the hospitalization likelihood of each individual patient and provide extra monitoring or even change in the therapy plan.
In this project, we use the patients’ CT images before the surgery to predict LVMI regression. The fully automatic process starts with a tool which captures the left ventricle geometry from CT images. We then apply several machine learning processes to analyze the shapes. This statistical approach explores shapes differences and similarities from the population average and summarizes shapes’ information as representative shape numbers. Finally, LVMI regression is estimated through machine learning predictor algorithms correlating the shape numbers with recorded follow up data. Our results demonstrate tons of information hidden in the medical images which could be leveraged towards better diagnosis and therapy planning.