June 2018 - Present: Evaluating the Performance of Deep Learning and Traditional Radiomics Algorithms on the Lung Nodule Classification Task in 3-D Thoracic CT Scans
Mentors: Dr. Daniel Rubin and Dr. Blaine Rister, Stanford University

Many techniques, ranging from convolutional neural networks (CNNs) to traditional radiomics, are used for classification of tumors. The goal of this project is to compare the performances of CNNs and traditional radiomics techniques in similar conditions and determine if and when one method achieves higher levels of performance than another. 

​

Here is a link to a preprint for a paper that I am a co-author on.

​

June 2017 - Present: A Weak Supervision Machine Learning Approach to Classification Using an XGBoost Model
Mentor: Dr. Stephen Bach, Brown University (previous Stanford University postdoc) and Dr. Christopher Re (Stanford University)

In many situations, it may be difficult to acquire labels for a large set of data in order to train a conventional machine learning (ML) model. A technique called weak supervision combines employs rule-based labeling functions to automatically but noisily label training data, which can then be trained on using a conventional ML model. The goal of the project is to determine whether the weak supervision framework (specifically Snorkel, being developed by Stanford researchers) will work well for classification problems when using XGBoost, a recently popular gradient boosting algorithm, for the discriminative model. In sum, Snorkel + XGBoost performs extremely well on classification problems where writing labeling functions is intuitive for humans. Serving as the discriminative model, XGBoost performs well and yields high scores, but the relationship between the performances of XGBoost and other algorithms is still being investigated. The weak supervision techniques explored in this study can save companies millions of dollars in hiring hand-labelers and enable them to utilize unlabeled data by hiring just a few domain experts who can write rule-based labeling functions.

​

​

​

September 2016 - March 2017: A Machine-Learning Approach to Correlate Environmental and Demographic Factors to Cancer Incidences Across US Counties

Independent Project

This study seeks to identify environmental and demographic factors contributing to cancer incidences, approximately 80% of which are classified as sporadic, across the United States. All data utilized was public, containing a variety of statistics by U.S. county. Through the use of machine learning and statistical techniques, an average of 20 demographic and chemical factors are found to be strongly associated with increased or decreased cancer rates, across the four different cancer types examined. Even though not all of these factors may be causes, they have strong correlations with the cancer incidence rate, and thus deeper examination of them can reveal the true risk factors. Thus, this study represents a novel way of bringing to light various previously unexamined factors that could be causing cancer incidences and therefore could represent a major step forward in preventive medicine.

 

 

 

 

 

 

​

 

 

 

 

 

 

 

 

 

 

 

 

 

July 2015 - March 2016: The Effect of Simulated Microgravity on Tissue Regeneration and its Mitigation Through the Use of Resolvin D1

Mentor: Dr. Anuran Chatterjee, UCSF. Partner project.