This project extends the xMADD diffusion framework to generate clinically realistic synthetic chest X-ray and CT images conditioned on demographic and physiological factors such as age, sex, BMI, and disease state. Using large-scale MIDRC datasets, it focuses on building and optimizing conditional diffusion training pipelines with cross-attention-based multimodal conditioning. The work also develops evaluation methods to assess both image quality and phenotype fidelity, aiming to advance trustworthy generative modeling for healthcare and provide reproducible benchmarking tools for the research community.

Through this SuperUROP project, I am looking forward to experimenting with state-of-the-art diffusion models and multimodal learning in a biomedical setting. I hope to strengthen my skills in working with large-scale datasets and in building, benchmarking, and optimizing generative model training pipelines in a research environment.