I’m Udi. I build at the intersection of physics, medicine, and software.

Medical Physics PhD student at UW–Madison, building research tools around imaging, AI, and radiation oncology.

Udbhav Ram outside the McMaster physics building
Interactive synthetic volume

Move through a CT slice.

This small, code-generated phantom shows how a stack of cross-sections becomes a volume. It uses no patient information and is not a clinical image or model result.

See the related segmentation research
Synthetic anatomy phantomSlice 49 / 96
Not patient data
SYNTHETIC PHANTOM · FOR INTERACTION ONLY

Research

Recent work ranges from radiology decision models to treatment-planning and imaging studies.

Collage of MR imaging, code, CT imaging, and radiation dose distribution
October 13, 2026 · medRxiv preprint · submitting author

RadKev: a decision model built for radiology.

RadKev is an open-weight AI model for radiology decisions with a fixed set of answers. It can run inside an institution, returns probabilities in one step, and was evaluated against general decision models and language models across fifteen tasks.

AAPM poster describing a locally hosted language-model pipeline for TG-263 naming quality assurance
2025 · AAPM Blue Ribbon Poster

Teaching a model the clinic’s language.

A treatment-target name is a small piece of text with a large job: it has to preserve clinical intent while fitting a shared naming convention. This project explores where local language models can help, and where a clinician still needs to look closely.

Read the case study
Udbhav Ram presenting clinical AI research at UAB Radiation Oncology
2025 · First-author journal article · JACMP

A closer look at the treatment plan.

For people with multiple brain metastases, a radiation plan needs to cover the targets while limiting dose outside them. This study asks how high-fidelity planning and control rings change that balance on Ethos 2.0.

Read the case study
Published comparison of abdominal organ contours from nnU-Net, Auto3DSeg, SwinUNETR, and ground truth
2025 · First-author journal article · Intelligent Oncology

A contour has to work for a person.

Good overlap scores are useful, but they do not tell the whole story of an automatically drawn organ. This head-to-head study pairs quantitative evaluation with blinded physician review.

Read the case study
Arrow McLaren IndyCar race car on track
Arrow McLaren IndyCar · 2023
Outside the lab

Motorsport

In 2023, I worked at Arrow McLaren IndyCar as a data-and-strategy intern. I built race-strategy simulations and worked with telemetry and performance tools during race weekends. Formula SAE and test-driving taught me the same lesson from another angle: a model is useful only when it helps someone make a decision. I still make time for music, riding, and diving, but motorsport is the thread that most directly shapes how I build.

More beyond the lab →