Research & Projects
Automated Bias Reduction in Deep Learning Based Melanoma Diagnosis using a Semi-Supervised Algorithm
Regeneron ISEF Finalist, Global Health Leaders Conference @ JHU, IEEE BIBM, Virginia State Science Fair Grand Prize Winner
My goal in this research was to design an automated method to diagnose Melanoma, a dangerous form of skin cancer, directly from images of the skin. However, the main portion of my work was focused on making sure the machine learning model was not biased towards certain skin colors or general skin features. I used a model called the LatentNet to focus more on under-represented features while training the model which resulted in a more generalizable model. I had the opportunity to present this research at an IEEE Conference and at the 2021 Regeneron ISEF competition.
Optimizing Prediction of MGMT Promoter Methylation from MRI Scans using Adversarial Learning
Accepted to 34th IEEE ICTAI Conference, Global Health Leaders Conference @ JHU, Presented to CAIDM Group @ UC Irvine
In this research, I propose a new method for coregistering MRI scans by using a model called the Intermediate State Generator to generate synthetic MRI slices. This could potentially increase the amount of information presented in a scan. I used this while building an automated system to determine the MGMT Promoter Methylation status directly from MRI scans of Glioblastoma patients. Determining this molecular feature plays a key role in deciding whether to provide a patient with chemotherapy treatment. I am currently pursuing this project further with the CAIDM lab at UC Irvine.
Cross-Dataset Evaluation of Multimodal Neural
Networks for Glaucoma Diagnosis
Presented at 9th IEEE DSAA Conference
I originally began researching the topic of AI-based Glaucoma diagnosis with the Aravind Eye Hospital in 10th grade during a short collaboration. With access to several medically annotated datasets, I briefly worked with a mentor and learned how to create segmentation and classification models with the data. About a year later with a team of six students, I wrote a paper that placed emphasis on creating a generalizable model. We measured generalizability using a method called “Cross-Dataset Evaluation”. Our model utilized incorporated Convolutional Neural Networks along with mathematically extracted features such as the Cup-to-Disc Ratio and Eccentricities. Our paper was accepted after peer-review and I presented this research at the 9th International Conference of Data Science and Advanced Analytics.
Ichos: A Web Application for Early Disease Screening Using Speech and Breath Recordings
Congressional App Challenge Winner, Invited to Volunteer at “Apply Yourself” hosted by USPTO, Project Displayed in US Capitol
We developed Ichos as a simple and efficient tool to diagnose a variety of diseases including Alzheimer’s disease, Specific Language Impairment, and a variety of other respiratory diseases. Each of these were diagnosed using an automated system that relied on an audio input including breath and voice recordings. We hope that this tool can help improve the accessibility of important diagnosis tools.
Our team was requested to speak at an event called “Apply Yourself” to help younger students design their own applications. This event was hosted in the Clara Barton Auditorium in Alexandria, VA.
Sentiment Sensitive Debiasing: A Learning-Based Approach to Remove Ethnic Stereotypes in Word Embeddings
Published in International Journal of Computational Linguistics (Co-Author)
Physical Education Tool (PET)
Best Educational Hack @ UVA HooHacks
PET is a software that can be used to help physical education teachers ensure their students are performing various exercises correctly through a virtual environment. The user uploads a video o them doing a squat, curl-up, bicep curl, plan, and/or pushup to the website, and PET provides real-time feedback of what, if anything, the user needs to improve about their form.