Research

Problem: Clinical image datasets are small and expensive to annotate, while the morphological variation that carries diagnostic signal — red cell shape, tissue morphology — is exactly what supervised models under-learn from a few hundred labels.
Approach: Pre-train representations from unlabelled images with contrastive objectives, and design the negative sampling so that visually similar but clinically different samples are the ones the model has to separate.
Work: SLRBC, SPIE Medical Imaging 2025

Problem: Brain connectivity is a graph, not a grid, so convolutional inductive biases fit it poorly. The structure that distinguishes clinical populations lives in the connections between regions, which a standard image encoder tends to flatten away.
Approach: Learn over the connectivity graph directly, combining a transformer encoder with self-supervised pretraining to make the most of small cohorts.
Work: Graph transformer for autism detection, IEEE ISBI 2025

Problem: Large-area watermarks destroy most of what a restoration model would otherwise condition on. Treating removal as plain image-to-image translation leaves residual traces, and methods that assume a pixel-perfect mask fail on the rough masks available in practice.
Approach: Separate watermark exclusion from background restoration, recover content both from what remains visible beneath a translucent watermark and from the surrounding context, and train so that coarse masks still work.
Work: RIRCI, AAAI 2024 · Inpainting priors for large-area removal, AAAI 2025