Research
My research interests are broad but generally can be categorized
as a fascination with neural systems.
My PhD thesis
(completed in 2021) was focused on designing deep learning algorithms for robust and efficient automated echocardiography
analysis. This included two algorithms which were implemented within GE Healthcare's most recent software
release (measurement and
classification). After recognizing the
difficulty of acquiring high quality labeled medical data, I also created
a framework for generating
synthetic data from anatomical models.
Prior to my PhD, I worked on developing brain inspired computing in hardware. I developed my own
neural network in hardware
and spent a year researching neuromorphic design with the Stanford Brains in Silicon lab. I also took
a break from academic research for a year to found
a start-up with some friends from the Stanford business
school. We built a consumer-facing product using machine learning to provide travel recommendation systems.
I spent my undergrad working on computational neuroscience simulations for brain machine interfaces with the
wireless nanosystems group at the University of Utah and an internship with the
IT'IS Foundation in Zurich. I did my bachelors thesis on
computational modeling of the rat hippocampus for improving neural stimulation.
Please see my Google Scholar
for a full list of my research publications.
Automatic echocardiography analysis using deep learning
Efficient measurement, workflow, and data generation
My PhD thesis conducted at the University of Oslo in partnership with GE Healthcare. You can download a pdf of my thesis or see my defense lecture slides. The slides for my trial lecture on strain and deformation estimation in cardiac estimation are also available.
Data generation for medical images
A framework for automatic generation of large, heterogeneous, synthetic datasets using generative adversarial networks. The framework replaces costly medical image acquisition and labeling and allows for deep learning algorithms to be trained from high quality data derived from anatomical models. For more information see:
- Publication in Transactions on Medical Imaging.
- Website describing the project in more detail.
- GitHub repository containing all code for the example with echocardiography segmentation.
Automatic left ventricle dimension measurement
A CNN-based system for automatically measuring the dimensions of the left ventricle from parasternal long axis echocardiography images with high accuracy. This project was implemented in the Vivid Ultra Edition software release from GE Healthcare as one of three key AI features. It automates the highly redundant, yet difficult, task of quantifying chamber size bringing significant time-savings to users. For more information see:
- The publication in the smart ultrasound imaging workshop at the 2019 MICCAI conference where it won the best paper award.
- The overview page or application demo from the Vivid software release.
Doppler spectrum classification
A CNN-based system for automatically classifying Doppler spectra from echocardiography images. This project was implemented in the Vivid Ultra Edition software release from GE Healthcare as one of three key AI features. It automates a key step in the measurement pipeline providing time-savings for users. For more information see:
- The publication in the Journal of Biomedical and Health Informatics.
- The overview page or application demo from the Vivid software release.
Other projects
My GitHub gives some examples of some other fun projects I've worked on, including a CNN-based markerless pose estimation, a collaborative filtering recommendation system, or deep learning for image de-blurring.
