profile

Hello! Currently I am a PhD candidate in applied mathematics at the University of Houston. I have the privilege of collaborating with Robert Azencott and Andreas Mang on automatic classification of medical images. My academic path has provided me with a strong foundation in mathematical modeling, numerical development, and machine learning algorithms. Through completing coursework projects, I acquired fundamental knowledge and hands-on experience with a diverse range of classification, clustering, and spatial regression models.

In our medical study supported by Houston Methodist, we tackled 3D image processing and disease classification problems. Handling MRI and echocardiography images equipped me with insights into the complexity of raw data, inspiring exploration and development of data denoising techniques and strategies to enhance model robustness. Along the journey, I mastered well-established machine learning frameworks such as TensorFlow, scikit-learn, and others. In addition, the demands of extensive research led me to implementation of more advanced and custom models and techniques. Dealing with large datasets also gained me expertise in cluster and cloud computing utilizing workload managers, parallelization, and GPU computing.

During my internship at Aikynetix as computer vision engineer I actively contributed to human pose and physical parameters estimation development. I got myself familiar with top-notch computer vision models, transfer learning, and various machine learning toolboxes such as mmlab, mediapipe, coreml, and so forth. I successfully worked within a team setting. My teamwork experiences have improved my communication and commitment to working together to reach common goals efficiently.

Now I am actively looking for opportunities in ML/AI and Data Science. Please feel free to contact me.


profile

Our team developed a method to predict cardiac diagnoses based on 3D images of the mitral valve using Diffeomorphic Registration (DR) and classical machine learning algorithms. DR is a powerful mathematical tool that we utilize to derive geometrically insightful metrics, such as strain intensities and kinetic energy. The strain intensities describe the level of deformations that surfaces undergo to match each other, while kinetic energy reflects the amount of effort required to register the matching surfaces. Both metrics provide meaningful shape features, which allow us to represent the data in a lower-dimensional space of distances. Please refer to the paper for more details on this work.


20elim

Selecting a subset of images or shapes that accurately represents the entire set is often a challenging problem. We propose a distance-based clustering method to identify a reference set that effectively captures the characteristics of the whole class with fewer shapes. This method offers versatile applications, including clone removal, improved visualization, and data compression. We apply optimal transport techniques to map shapes into a feature space and use constrained K-means algorithms to identify the most central images.