Research
Deep learning across imaging, video, genomics and clinical records. Special attention to what decides whether a model holds up outside its training set: multimodal fusion, calibrated uncertainty, and generalization across scanners, sites and capture conditions.
The new anti-amyloid drugs work, but a fraction of patients develop ARIA, swelling or bleeding in the brain, and that is the main thing limiting how widely the drugs get used. I built a multimodal model that estimates a patient's risk before treatment starts, combining imaging, genetics and clinical variables.
npj Dementia, in press
A tumor segmentation model will draw a boundary whether or not it has any business being confident about one. This work attaches calibrated uncertainty to the output, so the regions the model is guessing at are visible instead of hidden behind a clean outline.
Primary CNS lymphoma carries genomic markers that normally require a biopsy to find. We trained interpretable models to predict some of them from routine clinical brain MRI instead, which is the sort of thing that only matters if a clinician can see why the model said it.
ASFNR, NCCN
At CurieDx I build the models behind point-of-care screening: smartphone video of a throat, fused with structured symptoms. The images are captured by patients rather than clinicians, so most of the real work is distribution shift, validation across sites, and calibration, not architecture.
NIH NIBIB, BARDA
Deformable registration between intraoperative ultrasound and pre-operative CT, so a surgeon gets spatial context without leaning as hard on fluoroscopy. Robotics, registration, segmentation and interface design all had to hold up under real clinical constraints at once.
Adapting a pretrained vision model to spot dislocations across hips, shoulders, elbows, knees and arthroplasty cases. The interesting constraint was never accuracy on one dataset, it was staying usable across very heterogeneous imaging.