Medical AI Research
First-author doctoral research integrating bone-marrow single-cell morphology with clinical variables to predict overall survival in newly diagnosed multiple myeloma.
Evidence mode
Cancer prognosis depends on both microscopic morphology and patient-level clinical variables. The technical challenge is to convert millions of single-cell images into patient-level survival signals without losing clinically relevant heterogeneity.
Designed an end-to-end multimodal pipeline that uses self-supervised cell embeddings, a malignancy index, multiple-instance survival modeling, and late fusion with clinical Cox models.
Built Python/PyTorch feature pipelines with HDF5 storage, 384-dimensional image embeddings, patient-level bag construction, survival metrics, and interpretable risk stratification.
Improved C-index from 0.73 for clinical-only modeling to 0.76 after image-clinical fusion; time-dependent AUROC reached up to 0.84 and high-vs-low risk stratification showed HR around 4.0 with p < 0.001.