Research & Selected Publications

All publications on Google Scholar · * Equal contribution

Latent diffusion model trained on unlabeled NLST and TCGA pathology images, conditioned on cohort, tissue preparation, and UNI features.

2026 · IEEE Journal of Biomedical and Health Informatics

Semi-Supervised Domain Adaptation with Latent Diffusion for Pathology Image Classification

Tengyue Zhang, Ruiwen Ding, Luoting Zhuang, Yuxiao Wu, Erika F. Rodriguez, William Hsu

We use a latent diffusion model trained on unlabeled source and target images to generate pathology images that preserve tissue morphology while incorporating target-domain appearance. Adding these synthetic images to classifier training improved lung adenocarcinoma growth-pattern classification on a held-out target cohort, increasing weighted F1 from 0.611 to 0.706.

DOI Code

Evaluation framework comparing partial information decomposition metrics with single-modality, early-fusion, and late-fusion model performance.

2025 · Journal of Biomedical Informatics, 167, 104833

Evaluating an information theoretic approach for selecting multimodal data fusion methods

Tengyue Zhang*, Ruiwen Ding*, Kha-Dinh Luong, William Hsu

Can information theory tell us how to combine medical data? We evaluated partial information decomposition across seven modality pairs from four biomedical cohorts. The metrics did not consistently identify the best-performing fusion strategy; we investigated parameter selection and uncertainty estimation to improve their reliability.

DOI Code

CT-derived radiomic features and whole-slide pathology features are selected and fused for progression-free survival prediction with Cox regression.

2024 · SPIE Medical Imaging: Digital and Computational Pathology

Using integrated radiomic and pathomic-based models to predict progression-free survival in early-stage lung adenocarcinoma

Tengyue Zhang*, Anil Yadav*, Ruiwen Ding*, Sean Johnson, Denise Aberle, Ashley Prosper, Erika Rodriguez, Ana Cristina Araujo Lemos da Silva, William Hsu

We combined features from CT scans and whole-slide pathology images to predict progression-free survival in 106 patients from the National Lung Screening Trial. With repeated five-fold cross-validation, the fused Cox model achieved a mean test C-index of 0.634, compared with 0.612 for radiomics alone and 0.584 for pathomics alone.

DOI

Medical images pass through preprocessing and the Segment Anything Model, followed by a probability map generator and level-set active contour refinement.

2024 · SPIE Medical Imaging: Computer-Aided Diagnosis

Refining Boundaries of the Segment Anything Model in Medical Images Using an Active Contour Model

Noor Nakhaei, Tengyue Zhang, Demetri Terzopoulos, William Hsu

We combine the Segment Anything Model with a level-set active contour model to refine medical image boundaries at test time. Evaluated on skin lesion and retinal vessel datasets, the approach addresses segmentation errors in images with low contrast or artifacts without additional supervised fine-tuning of SAM.

DOI