Research

Longitudinal medical imaging tracks how disease progresses over time, but forecasting future scans is difficult because clinically meaningful changes are spatially complex and often irregular in timing. This project developed a text-conditioned video diffusion model for temporal modeling of 2D medical imaging tasks: first pre-training on individual frames, then extending the architecture into a spatio-temporal U-Net that generates full image sequences, with masking to support padded, irregularly-spaced longitudinal inputs.

The framework was validated on a synthetic dataset built from a known linear time-invariant system before being applied to real clinical data, including ADNI brain MRI, longitudinal skin lesion images, and chest X-rays. MRI experiments produced realistic-looking sequences but also revealed the limits of a 2D slice-based model applied to inherently 3D anatomy, which motivated a shift toward lower-resolution 2D modalities that better match the model’s assumptions. On skin and chest X-ray data, the model was benchmarked against a ControlNet baseline using pixel-space error and CLIP/BiomedCLIP embedding similarity: history-guided sampling achieved the lowest pixel error on skin images and the strongest embedding similarity overall, while causal sampling performed best in pixel space for chest X-rays. The result is an end-to-end pipeline for temporal modeling of 2D medical imaging with video diffusion models.

Early cancer detection dramatically improves outcomes, but the sequencing needed to read methylation patterns in circulating cell-free DNA (cfDNA) is expensive and logistically demanding. This project asked whether accurate detection is still possible from a tiny, cheap sample. Using whole-genome bisulfite sequencing data from 239 breast cancer patients and 278 healthy subjects, I randomly subsampled just 1 million reads per sample to mimic the information available from a dried blood spot, then used a Mann-Whitney U test to select the most informative genomic regions and trained machine learning models to classify cancer versus non-cancer.

XGBoost reached an AUC of 0.91, outperforming Random Forest (0.87), a CNN (0.77), and an MLP (0.72). Notably, this matched the 0.91 AUC of a published multimodal model that combined genome-wide methylation, copy number, and fragment end-motif features, showing that position-specific methylation alone can be just as discriminative while needing far less data. Gene enrichment analysis of the top hypermethylated regions recovered known tumor-suppressor pathways (cadherin-mediated adhesion, Wnt and Hippo signaling), confirming the selected features were biologically meaningful. Together the results point toward low-cost, minimally invasive cancer screening that could reach beyond traditional clinical settings.