MICCAI 2026
Combines healthy-skin inpainting, rare-lesion mapping, and disease-conditioned diffusion to create 266k+ synthetic dermatology images across skin tones and improve malignancy classification.
90.9% DDI malignancy accuracy, with equality of opportunity improving from 69.6% to 86.6% versus PatchAlign after real-data fine-tuning.
MICCAI 2026
Adapts a monocular 3D model with synthetic D-Synth data and small clinical datasets to recover metric lesion geometry and detailed surface normals from one image.
Synthetic training reduces SKINL2 depth-scale overestimation from 16.1× to 1.11× while fine-tuning only 0.6% of model parameters.
Combines portable fluorescent imaging with spatial clustering to map pH and nitric oxide in mouse wounds and relate biomarker patterns to healing progression.
Across eight mouse wounds, smaller, brighter nitric-oxide regions correlate with later healing days and greater re-epithelialization.
arXiv · 2025Preprint
Trains a billion-parameter world model on street-view images linked by location, time, and movement commands, unifying controlled image generation, geolocation, and action prediction.
60% of geolocation predictions fall within 10 meters on STRIDE’s San Mateo test data.
Tests remotely controlled, multi-channel fluoxetine delivery with onboard wound imaging in a porcine model, demonstrating improved early repair relative to standard care.
At day 3, treated porcine wounds show 37% greater re-epithelialization and a 33% lower M1/M2 ratio than standard-care controls.
Demonstrates a standalone wearable ion pump that delivers scheduled fluoxetine doses, improving early tissue repair and inflammatory markers in a mouse-wound proof of concept.
At day 3, treated mouse wounds show 39.9% greater re-epithelialization and 27.2% lower M1/M2 ratios than inactive-device controls.
Uses a frozen diffusion model’s visual features with lightweight prediction heads to segment skin lesions and classify malignancy from small, skin-tone-balanced labeled datasets.
81% DDI malignancy accuracy from 60 labeled training images, with no diffusion-backbone fine-tuning.
Learns temporal ordering from mouse-wound image pairs, clusters the resulting features into four healing stages, and trains a classifier without manual stage annotations.
90.6% test accuracy against cluster-derived healing-stage labels on 32 held-out mouse-wound images.
Combines wound detection, segmentation, and reference-based sizing to automate mouse-wound measurements, preserving closure trends despite uneven imaging and missing or damaged splints.
Post-processing raises U-Net test Dice from 67.85% to 86.65% for segmenting mouse wounds and splints.
Studies honeybee re-identification from abdomen images, using short video tracks for self-supervised pretraining and tag-derived identity labels for fine-tuning across days.
62.4% rank-1 identification across days, choosing among 11 gallery images using abdomen appearance alone.