Publications

We present MamoCLIP, a federated contrastive learning framework for mammography analysis that leverages multi-institutional data without centralizing patient records. Our approach achieves state-of-the-art performance on EMBED v2, demonstrating robust cross-site generalization for breast density and lesion classification.

APIC extracts tumor-immune interaction features from standard H&E pathology slides to predict treatment benefit in prostate cancer patients. Validated on CHAARTED and STAMPEDE clinical trial datasets, APIC provides actionable prognostic stratification without the cost burden of molecular assays.

HistoQC provides automated quality assessment for whole-slide pathology images, detecting artifacts such as blur, fold, bubbles, and pen marks. With thousands of downloads, it has become the standard tool for preprocessing digital pathology slides in computational pathology pipelines.

Stain variability across institutions is a major challenge in computational pathology. F-SYN uses Fourier-domain normalization to harmonize pathology slides without introducing GAN-related artifacts, achieving superior structural preservation compared to deep learning-based stain transfer methods.

MOSCARD introduces a causal de-confounding framework for multimodal opportunistic screening, addressing spurious correlations that arise when fusing imaging with structured clinical data. The method demonstrates improved calibration and fairness across demographic subgroups.

We develop and validate a radiomics-based model for estimating prostate biological age from standard biparametric MRI, derived from the VA dataset of 387 patients and validated across Mayo Clinic biobank data. Biological age discordance is associated with biochemical recurrence after definitive therapy.

A comprehensive PRISMA-compliant systematic review of 87 studies examining multimodal fusion approaches for breast cancer prognosis. We identify key methodological gaps in handling missing modalities, lack of diverse cohort validation, and limited clinical interpretability — gaps that MEFINDER directly addresses.

We present a scalable NLP toolkit for extracting ER-positive breast cancer recurrence labels from clinical notes in EMBED v2 (260,815 patients). The toolkit achieves 91% F1 score against manually annotated gold labels and enables large-scale outcome curation without manual chart review.