Generative AI

Overview

Generating and modeling dynamic scientific data for analysis, education, visualization, and privacy-aware data sharing. This line of work produced a peer-reviewed publication in JAMA Ophthalmology (2025).

Research Problem

Medical data sharing is constrained by privacy: real patient imaging cannot circulate freely, yet research, education, and model development all need realistic data. The question was whether a generative model could produce clinically meaningful dynamic imaging from text descriptions — enabling sharing without exposing patients.

Method

A text-to-video model was developed to generate dynamic fundus fluorescein angiography (FFA) videos from clinical text descriptions, built with Python and PyTorch.

Model Architecture

The model combines a Wavelet-Flow variational autoencoder (Wavelet-Flow VAE) with a Diffusion Transformer, enabling spatiotemporal modeling and visual generation of fundus lesion features.

Dataset

The model was trained and evaluated on 3,625 anonymized FFA videos.

Evaluation

Generation quality and privacy protection were assessed through a combination of:

• Objective metrics
• Text–video semantic consistency
• Subjective evaluation by ophthalmologists
• Image retrieval metrics

Personal Contribution

I am a co-first author of the paper. I contributed to study conception and design, data acquisition and interpretation, manuscript drafting and statistical analysis, as well as model development, including the integration of the Wavelet-Flow VAE with the Diffusion Transformer for spatiotemporal modeling.

Publication

Generation of Fundus Fluorescein Angiography Videos for Health Care Data Sharing. JAMA Ophthalmology. 2025; 143(8): 623–632. DOI: 10.1001/jamaophthalmol.2025.1419

Ethics and Privacy Context

The work is motivated by privacy-aware data sharing: training and evaluation used anonymized data, and the generated videos are intended for clinical research, medical education, and privacy-preserving data sharing scenarios.