Research Interests

Synthetic axial brain MRI illustration in a FLAIR-like contrast: a large bright lesion in the left half of the image and a smaller, fainter bright spot in the right half, beside the lateral ventricles. R1 R2 Initial Reviewed

01 · Learn

Learning from Imperfect Supervision

Can useful medical AI emerge from sparse, incomplete expert annotations?

Clinician · Sparse annotation

Opens the scan and marks only what is quick to mark. A short scribble on the prominent lesion (R1). No full outline is drawn.

Promptable Foundation Model

Awaiting a prompt Prompt received · producing an initial mask Initial mask covers R1 only

Potential omission

Uncertainty is elevated near a smaller region (R2). With limited supervision, an initial mask can be incomplete.

02 · Collaborate

Decisions That Can Be Challenged

What happens when a clinician disagrees with an AI decision?

Clinician · Challenge on R2

“Could the smaller lesion have been missed?”

Image Evidence Possible unsegmented abnormality near the marked region.
Consistency Review The current mask does not cover the challenged region.

Evidence synthesis

Challenge supported for further review

Proposed revision: inspect and include the smaller region

Two independent checks justify a review. Agreement is not clinical proof.

Clinician reviews the proposed correction

03 · Improve

Systems That Grow With Their Experts

How can verified corrections become reusable knowledge?

Segmentation revision

Initial mask · R1 Reviewed revision · R1 + R2

Verified Feedback Memory

Issue
Potential lesion omission
Evidence
Image region + expert objection
Resolution
Segmentation revised after review
Status
Verified in demonstration

Retrieved as guidance for future cases. It does not rewrite the model's weights.

Verified feedback becomes future guidance

Learn from experts. Reason with experts. Improve through experts.

From Prediction to Partnership

When a wrong answer is costly, accuracy alone is not enough.

Most AI is trained once on carefully labeled data, then deployed and left unchanged. Medicine does not work that way: expert time is scarce, clinicians remain accountable for every decision, and practice keeps changing. I want AI that works alongside experts in this setting, and three connected questions guide my work.

  1. Learn

    Learning from what experts can actually give

    How can AI learn reliably when supervision is scarce, imperfect, and never complete?

    Expert attention is the scarcest resource in medicine, so the future of medical AI cannot depend on ever-larger labeled datasets. I study learning that takes supervision as it really is: sparse, noisy, and uneven across classes and hospitals. The goal is models that know how much to trust each signal they learn from, and that state their uncertainty honestly when the evidence runs out.

    A model trained on imperfect evidence will sometimes be wrong, so it must be open to challenge.

  2. Collaborate

    Decisions that can be challenged

    How can people not only understand an AI decision, but contest and correct it?

    Explanation is a one-way street: the system speaks and the human listens. Real oversight needs the other direction. I work toward contestable AI, where reasoning is laid out as explicit evidence and argument, any step can be disputed, and a valid objection changes the outcome. As decisions move to societies of AI agents, the same standard must hold among the agents themselves: agreement should reflect independent evidence, not repetition or persuasion.

    Every challenge an expert raises is knowledge the system did not have.

  3. Improve

    Systems that grow with their experts

    How can everyday human feedback become lasting improvement?

    Most deployed models are frozen at training time, and the corrections experts make every day are lost. I aim for self-improving systems that turn feedback, whether a correction, an objection, or a preference, into durable knowledge, safely and without costly retraining. This closes the loop: feedback becomes the next source of supervision, and the partnership grows stronger with use.