MITRP undergraduate researcher Jake Yun has been awarded a prestigious 2026-2027 Wisconsin Hilldale Undergraduate Research Fellowship. This competitive award provides Jake with a stipend to support his independent research over the upcoming academic year under the mentorship of MITRP Principal Investigator Dr. Alan McMillan.

Jake’s project takes aim at a massive hurdle in modern medical data science: the computational weight of artificial intelligence. While Vision Foundation Models (VFMs) are incredibly capable at identifying anatomical structures and disease markers, training these massive networks end-to-end for specific clinical tasks is expensive and requires immense computing power. Jake is exploring a highly efficient alternative called embedding-based learning. Instead of retraining a massive model from scratch, this method extracts the model’s core “understanding” as a compact mathematical representation, which can then be used to train lightweight clinical tools on standard commodity hardware like a laptop. The problem is that researchers currently lack a standardized way to compare which foundation models actually perform best in these low-compute scenarios.

To bridge this gap, Jake is Helping to build MedRankAI, an automated, model-agnostic benchmarking pipeline designed to systematically test and rank foundation models. MedRankAI serves as a standardized proving ground, running AI models through a suite of diagnostic tasks—like classification, similarity retrieval, and unsupervised clustering—to see how well they process complex medical data. The pipeline will test models against a diverse array of public datasets covering dermatology, mammography, and radiography, while also evaluating them on anonymized, high-resolution CT and MRI sequences from the UW-Madison Department of Radiology. Developing new ways to extract reliable structural and functional information from advanced imaging scans is the core of what we do at MITRP, and Jake’s framework ensures that the AI tools we use to analyze those scans are evaluated rigorously for potential biases before ever touching real patient data.

By automating the testing process and eliminating the need for costly full-model training, MedRankAI will produce transparent leaderboards that highlight the trade-offs between a model’s clinical accuracy and its computational efficiency. This will ultimately help healthcare providers and researchers easily select the most cost-effective, reliable AI tools for their specific diagnostic workflows.

Congrats to Jake on a job well done!