Profile
- History of successful work in interdisciplinary & intercultural environments.
- Motivated to leverage deep understanding of machine learning algorithms and workflows to create a positive impact.
- Passion for creating machine learning applications, ensuring the highest standards of quality, as well as scientific rigor.
- Creative, collaborative, and innovation focused.
Experience
Data-Efficient Deep Learning for Biomarker Prediction in Angiosarcoma
- Development of an open-source pipeline (PEERCE) leveraging pre-trained generalist models and fine-tuned segmentation/classification networks for PD-L1 Tumor Proportion Score (TPS) prediction.
- Validated the tool’s utility to improve pathologist concordance. Published in Journal of Pathology Informatics.
Unsupervised Domain Adaptation for Generalist Cell Segmentation Models
- Design and implementation of SelfAdapt, a source-free Unsupervised Domain Adaptation (UDA) framework.
- Introduced and evaluated novel label-free early stopping criteria. Delivered the method as an open-source extension to the Cellpose ecosystem. Published in ICCV BIC.
PhenoBench: A Comprehensive Benchmark for Foundation Model Evaluation in Pathology
- Development of a benchmarking framework to evaluate the generalization of pathology foundation models under technical and medical domain shifts. Implemented in Python using PyTorch. Published in MICCAI.
Deep Learning for Tau PET Prediction in Alzheimer’s Disease
- Led a collaboration with Eli Lilly to develop an AI framework for predicting future Tau accumulation from baseline scans + clinical/genetic covariates (deep CNN feature extraction + tree-based models).
- Demonstrated superior performance in identifying rapid progressors, optimizing patient selection strategies.
Systematic benchmark to analyze CNN performance as well as relative generalization capabilities
- Use of simulated static patterns to assess CNN performance. CNN training and testing coded in Python using PyTorch. Experiments were run on a Stanford high performance cluster, as well as Google Cloud. Published.
Deep Learning to Predict Standardized Uptake Value Ratio and Amyloid Status on 18F-florbetapir
- Creation of deep learning pre-processing pipeline and CNN training on compute cluster. Published.
Predicting Future Imaging Biomarkers with Machine Learning: An Amyloid Study
- Employment of CNNs and gradient boosted decision tree to predict development of standardized uptake value ratio. Method was patented. Published.
Lead development and deployment of service for visual detection of erroneous production parts.
- Combined neural network classification with low-latency algorithmic image segmentation to identify manufacturing errors in production parts. Researched ways to reduce generalization error.
Implementation and benchmarking of object detecting CNNs to identify weeds in farming environments.
- Established new data preprocessing pipeline and modified TensorFlow Object Detection API for enhanced validation.
Use of big industry-specific dataset to create pretrained CNN weights for transfer learning. Resulting CNN presented to Bosch CEO at research presentation session.
- Modification of C++ Caffe Deep Learning Framework to allow for 4-channel CNN input. Implementation of improved data pre-processing pipeline.
Patents & Certifications
Co-inventor with Greg Zaharchuk. Methods for predicting biomarker progression in medical imaging to identify fast/moderate/slow progressors for clinical trial enrollment.
Stanford University Machine Learning Certificate via Coursera
Open Source Projects
Key open-source contributions in computational pathology and machine learning.
Publications
Selected peer-reviewed publications in machine learning, medical imaging, and computational pathology.