Roya Parsa

Roya Parsa

I teach machines to read tissue.

I'm a computer science master's student at Dartmouth, working in Dr. Saeed Hassanpour's lab on machine learning for digital pathology. I came into this through precision oncology, building proton-therapy tools where the model output guided a real treatment plan. That's why I care as much about whether a model is fair and usable as whether it scores well.

Open to research and industry roles in ML and bioinformatics

How I work

As a researcher

I adapt large-scale foundation models to massive, complex datasets. Emphasis is placed on interpretability and validation, so I build the tools necessary to understand what the model is learning and how it generalizes.

As an engineer

A model is only as good as the infrastructure supporting it. I engineer the whole pipeline. From data ingestion to compute scheduling and validation, I make sure the model can be trained and deployed reliably.

As a product thinker

I begin from the decision someone is trying to make, scope what the model has to support, and iterate until it's usable.

Selected work

Liver disease classification with Prov-GigaPath

Classified cholestatic liver disease from whole-slide images using GigaPath. Developed a multimodal fusion model that outperformed single-modality models by 12% in F1 score.

Paper in preparation

SPR-NET

A full-stack tool built with the New York Proton Center that reconstructs proton stopping-power maps from dual-energy CT, holding under 1% error. My U-Net variant beat analytic methods by 60% in high-density regions.

Paper submitted Repository

Do fairness fixes transfer?

Shortcut-mitigation methods are usually validated on one dataset and one population. I tested whether they hold up across Fitzpatrick skin types using Fitzpatrick17k and DDI — and whether the method rankings survive the move.

Paper in preparation

How robust are medical VLMs?

How do MedGemma's medical reasoning performance change as CT image quality is progressively degraded? Can evidence-grounded prompting improve robustness?

Paper in preparation

Smaller things — simpleCoder, a terminal coding agent with semantic RAG and file-permission guardrails, among others — live on GitHub.

Where I've worked

2026 - Present

AI Research Fellow

Imageomics Institute, The Ohio State University

A funded year of bioinformatics and imageomics research, including field work in Hawaii applying machine learning to environmental questions.

2025 - Present

Graduate Research Assistant

Hassanpour Lab, Geisel School of Medicine, Dartmouth

Computational pathology on whole-slide images, plus co-leading a YOLOv8 ultrasound system for cervical cancer screening in low-resource settings.

2024 - 2025

Student Researcher

New York Proton Center

Built CT analysis tools generating stopping-power maps under 1% error for proton treatment planning, and benchmarked dual-energy against single-energy models.

2023 - 2024

Software Engineer Intern

North Atlantic Industries, Bohemia, NY

Shipped a GPT-powered document search system that cut support tickets in half, and built Python ETL pipelines and Dockerized cross-compilation builds on AWS.

How I work

As a researcher

As an engineer

As a product thinker

Selected work

Liver disease classification with Prov-GigaPath

SPR-NET

Do fairness fixes transfer?

How robust are medical VLMs?

Where I've worked

AI Research Fellow

Graduate Research Assistant

Student Researcher

Software Engineer Intern

Say hello.