Seminar May 23 TH, 2023

Luigi E Perotti

Title: Combining MRI and computational models: from cardiac microstructure to kinematics measures of cardiac function.

Abstract: Computational models may help analyzing medical images of the heart to extract useful biomarkers of cardiac health. In this work, we present biomarkers related to the heart microstructure and motion, and how these biomarkers can be computed from cardiac diffusion tensor images (cDTI) and Displacement ENcoding with Stimulated Echoes (DENSE) MRI data.

First, we investigate how cDTI-based diffusion tensor measures may be employed to identify changes of the heart microstructure. In this context, we propose radial diffusivity as a biomarker of cardiac microstructural remodeling and interpret its changes using particle diffusion simulations. Subsequently we compute radial diffusivity, standard diffusion tensor invariants (such as mean diffusivity and fractional anisotropy), and contrast-based measures in swine subjects with myocardial infarction using both in vivo and ex vivo MRI data. Amongst all the non-contrast-based biomarkers, radial diffusivity increases significantly from remote to infarcted myocardium, while being more robust than changes in separate diffusion tensor eigenvalues.

Subsequently, we combine cardiac microstructure (cDTI) and motion (DENSE) data to evaluate myofiber strain Eff, a measure of cardiac deformation that is microstructurally anchored – i.e., independent of arbitrary reference systems – and directly linked to cardiomyocytes contraction and relaxation. We first estimate Eff by directly interpolating and differentiating Lagrangian voxel-wise myocardial displacements to compute the Green-Lagrangian strain tensor and then projecting it in the myofiber direction computed from cDTI. Subsequently, we derive a new method to evaluate Eff by minimizing the differences between displacements measured via DENSE and computed via a kinematics computational model. This second approach sidesteps the need to directly differentiate the measured displacement field and therefore limits the impact of experimental noise. 

Radial diffusivity and Eff represent two examples of biomarkers that could improve our evaluation of cardiac function. Ultimately, MRI informed computational models have the potential to provide new indices of cardiac function and help improving the diagnosis and prognosis for patients affected by cardiac diseases

Bio: Luigi Perotti received his Laurea (B.S./M.S.) degree in Civil Engineering from Politecnico di Milano, Italy, in 2004. Subsequently he continued his studies in Mechanical Engineering at Caltech where he received his M.S. in 2006 and his Ph.D. with a minor in Applied and Computational Mathematics in 2011. At the end of 2011, he joined Dr. Klug’s group in the Mechanical and Aerospace Engineering department at UCLA to pursue his research interests in biomechanics. Since then he has worked on several multidisciplinary projects involving collaborations across the Departments of Physics, Radiology, and the School of Medicine. In 2014 he received an AHA postdoctoral fellowship and joined Dr. Ennis' group in the Radiological Sciences department at UCLA. In 2017 he received an NIH K25 award to continue his research on combining computational models with MRI data and conduct pre-clinical studies. Dr. Perotti joined the Mechanical and Aerospace Engineering department at the University of Central Florida in 2019 where he leads the Computational Biomechanics Lab.

 Jonathan Araya Ugalde

Title: A Framework for Trustworthy Machine Learning Algorithm implementation for better decision making.

Abstract: The expansion of Machine Learning in many aspects of our daily lives is undeniable. The great versatility and predictive capacity of Machine Learning algorithms have led to their use in web advertising, facial recognition, medical diagnoses, among other fields. However, the emphasis on improving only the performance of these algorithms has neglected other aspects such as fairness or interpretability, which are important in decision-making contexts. It has been shown that Machine Learning algorithms reproduce the bias present in the data, which leads to discrimination (unequal treatment, usually harmful) against certain groups of people based on sensitive characteristics (such as gender or skin color). Furthermore, the most sophisticated Machine Learning algorithms operate as black boxes, whose functioning does not allow us to understand the attributes or decision rules that explain their predictions. In this regard, the problem of lack of trust in Machine Learning algorithms arises as a result of biased or non-interpretable predictions.

In this talk, the progress in the development of a framework that allows for the incorporation of techniques for bias reduction and explanation generation for a classification model predictions is presented, with the aim of providing a mechanism to quantify "confidence" in the predictions that can be contrasted with the confidence perceived by users of such models. The expected results aim to provide a tool that improves the level of confidence for users of Machine Learning models (both developers and users of ML-based systems) and promotes their use in decision-making contexts.

Bio: Jonathan Araya Ugalde obtained his degree in Computer Engineering from the University of Valparaíso (Chile) in 2016. In 2020, he began his PhD in Applied Informatics at the University of Valparaíso. His work focuses on fairness and interpretability in machine learning models.

Galería