PhD Research
Introduction
My PhD research focused on foundation models in medicine, emphasizing radiology and artificial intelligence. The cornerstone of my work was my dissertation project, which introduced the concept of comprehensive radiology AI. Additionally, I pursued projects detailed in the Other Research section.
Dissertation Research: Comprehensive Radiology AI
In my dissertation, I introduced comprehensive radiology AI as a holistic approach to AI-assisted medical image interpretation. This approach integrates diverse data types—including images, text, and structured data—to generate outputs evaluated using reproducible and automated performance metrics. My methods aim to advance radiology AI beyond bespoke solutions, enabling the development of generalist AI models for medical imaging.
Opportunistic Imaging
Central to comprehensive radiology AI is leveraging the full potential of medical imaging data beyond existing diagnostic workflows. This includes using images for tasks such as predicting future diseases or assessing biomarkers at a population scale to identify new insights. I developed AI methods to extract additional value from routine imaging data, including:
-
Ischemic Heart Disease Prediction: Developed the first multimodal methods leveraging abdominal CT and medical records to assess ischemic heart disease risk. Our fusion model outperformed clinical guideline-recommended models, achieving a 19% F1 score improvement. This approach utilizes routine images, offering additional diagnostic value without increasing radiation exposure. As a side bonus I also developed tissue saliency in this work, introducing the first quantitative explainability metrics for imaging classifiers at the tissue level. Read the paper.
-
Population-Wide Biomarker Analysis: Scaled methods to analyze imaging biomarkers across populations, uncovering hundreds of novel biomarker-disease associations for muscle biomarkers. Learn more at the Muscle PheWAS project page.
Radiology AI Evaluation
Comprehensive radiology AI requires robust evaluation methods. I introduced two key evaluation frameworks to assess AI model performance in radiology:
-
Improved Text Representations: Advanced radiology natural language processing (NLP) by creating a benchmark to measure performance in this domain. Findings were showcased at NeurIPS 2023.
-
CheXprompt: Developed and validated a GPT-4 based model for assessing generated radiology reports. The model aligns with human radiologist evaluations of quality, unlike prior metrics. CheXprompt was introduced in the LLaVA-Rad paper, utilizing large language model-curated data for state-of-the-art multimodal models for Chest X-ray report generation. Manuscript under review. Read the preprint. Use CheXprompt.
Other Research
In addition to my dissertation work, I collaborated on projects across various domains, expanding AI applications in medicine, including:
-
Automatic Radiology Protocoling: Developed tools for automating radiology protocoling to reduce time and errors. Findings shared in an IEEE paper.
-
Chest X-ray Generation: Contributed to creating RoentGen, a vision-language foundation model for chest X-ray generation. Paper at Nature Biomedical Engineering.
-
Multitask Disease Prediction: Designed a framework for chronic disease prediction as a multitask problem by training a single model for multiple diseases, improving label efficiency and performance. Demonstrated for cardiometabolic diseases using CT imaging (MICCAI 2022).
-
AI for Helping with Underdiagnosis: Demonstrated AI's use in bridging documentation gaps by diagnosing sarcopenia (low muscle mass) using CT imaging biomarkers, highlighting a significant documentation gap in medical records (<1% in records vs. ~30% prevalence in our patient population). Manuscript in press.
Acknowledgements
I was co-advised by Daniel Rubin, Akshay Chaudhari, and Curtis Langlotz. I also worked alongside other talented researchers in Stanford Radiology, including Robert Boutin and Andreas Loening, as well as external collaborators such as Leon Lenchik, Imon Banerjee and Bhavik Patel.