Embedded Files
About
About
I'm a Ph.D candidate in the Biomedical Engineering Department at Johns Hopkins University. I am honored to be co-advised by Prof. Jeremias Sulam and Prof. Alexander S. Popel. I am affiliated with the Mathematical Institute for Data Science (MINDs) and the Center for Imaging Science (CIS).
My research interest lies at the intersection of AI and healthcare, with a focus on computational pathology. I develop, improve, and adapt the state-of-the-art AI methods for digital pathology, inspired by actual clinical needs and practical computational challenges. Recent works include
enhancing image and annotation quality in histopathology images,
discovering prognostic biomarkers in spatial omics data,
explaining pathology foundation models with human-interpretable concepts.
I am fortunate to collaborate with colleagues across disciplines, including clinicians, pathologists, and industry partners. I'm passionate about translating research into scalable AI solutions to make real-world impacts.
Email: zwang218-at-jhu-dot-edu
Publications
Publications
Bi-level graph learning unveils prognosis-relevant tumor microenvironment patterns in breast multiplexed digital pathology
Bi-level graph learning unveils prognosis-relevant tumor microenvironment patterns in breast multiplexed digital pathology
Zhenzhen Wang ∙ Cesar A. Santa-Maria ∙ Aleksander S. Popel ∙ Jeremias Sulam
Patterns, 2025 (Cover article)
TL;DR: We introduced BiGraph, an unsupervised, multi-scale graph learning method for spatial single-cell analysis. It automatically discovers recurring cell patterns in the tumor microenvironment, associates them with survival outcomes, and nominates new biomarkers.
[paper] [code] [on the media]
Concept Bottleneck Model with Zero Performance Loss
Concept Bottleneck Model with Zero Performance Loss
Zhenzhen Wang ∙ Aleksander S. Popel ∙ Jeremias Sulam
The Second Conference on Parsimony and Learning (CPAL), 2025
TL;DR: We introduced CBM-zero, a method that converts any standard black-box models to concept bottleneck models (CBM). It does not comprise any prediction power, while provides concept-based explanations and enables human intervention.
Label cleaning multiple instance learning: Refining coarse annotations on single whole-slide images
Label cleaning multiple instance learning: Refining coarse annotations on single whole-slide images
Zhenzhen Wang; Carla Saoud; Sintawat Wangsiricharoen; Aaron W. James; Aleksander S. Popel; Jeremias Sulam
IEEE Transactions on Medical Imaging, 2022
TL;DR: We introduced a weakly-supervised method to automatically refine pathologists' coarse and imperfect annotations on Whole slide images. No extra data needed.
The use of weakly supervised machine learning for necrosis assessment in patients with osteosarcoma: A pilot study
The use of weakly supervised machine learning for necrosis assessment in patients with osteosarcoma: A pilot study
Christa L. LiBrizzi*, Zhenzhen Wang*, Jeremias Sulam, Aaron W. James, Adam S. Levin, Carol D. Morris
Journal of Orthopaedic Research, 2023
TL;DR: We applied and extended the LC-MIL method to osteosarcoma patients to estimate post-chemotherapy percentage necrosis. The model's estimation achieves a high correlation with the experts' estimation.
[paper] [code] [on the media]
Adaptive Explanations via Direct Preference Optimization
Adaptive Explanations via Direct Preference Optimization
Jacopo Teneggi · Zhenzhen Wang · Paul H. Yi · Tianmin Shu · Jeremias Sulam
Proceedings of the AAAI Symposium Series , 2025
TL;DR: We introduced a method that not only explains a black-box model's prediction with human-interpretable concepts, but also adapts explanations to different listeners' preferences. For example, while a medical doctor might understand an explanation in terms of clinical markers, a patient may prefer simple, layman's language.
PrivShap: A Finer-granularity Network Linearization
PrivShap: A Finer-granularity Network Linearization
Method for Private Inference
Method for Private Inference
Xiangrui Xu, Zhenzhen Wang, Rui Ning, Chunsheng Xin, Hongyi Wu
Transaction on Machine Learning Research (TMLR), 2025
TL;DR: We introduced a method that not only explains a black-box model's prediction with human-interpretable concepts, but also adapts explanations to different listeners' preferences. For example, while a medical doctor might understand an explanation in terms of clinical markers, a patient may prefer simple, layman's language.
[paper]
Page updated
Google Sites
Report abuse