Research
Most of the images I care about are ones a model has no business being confident about — a hundred labelled X-rays instead of a hundred thousand, microscopy of organisms nobody has catalogued, training sets where the easiest signal to learn is the wrong one. The four threads below are different angles on the same problem.
Medical imaging
Clinical images are few, acquired on different machines under different protocols, and a model that fails quietly on them is worse than no model at all.
We work on anatomical landmark detection in X-ray and MRI, lesion segmentation, and correcting motion artefacts without erasing the pathology underneath. Much of it is built on diffusion pre-training, so that a few dozen annotated scans are enough.
- WACV 2027 PACER: Curriculum Self-Training with Spatial Uncertainty for Pseudo-Label Selection in Few-Shot Anatomical Landmark Detection
- ISBI 2026 A Pair-Weighing Strategy for Enhancing Clip Zero-Shot Classification for Chest X-Rays
- ISBI 2026 Automated Landmark Detection for assessing hip conditions: A Cross-Modality Validation of MRI versus X-ray
- MICCAI 2026 CDPM-Align: Multi-Scale Guidance-Aligned Diffusion Pretraining for Robust Few-Shot Anatomical Landmark Detection
- BMC bioinformatics 2026 FiloAnalyzer: a deep learning approach for cell filopodia segmentation
21 papers in this thread — see all publications
Biological and microscopy imaging
Biologists photograph far more organisms than anyone can label, and the interesting ones are usually the ones no classifier has seen before.
We build classifiers for plankton, cells and fungal pathogens that work without large annotated sets, and anomaly detection that flags organisms falling outside the known classes rather than forcing them into one.
- BMC bioinformatics 2026 FiloAnalyzer: a deep learning approach for cell filopodia segmentation
- IJCNN 2025 Disentangled representations of microscopy images
- WACV 2025 In-domain self-supervised learning for plankton image classification on a budget
- Cell 2025 Phenotypic landscape of an invasive fungal pathogen reveals its unique biology
- Frontiers in Marine Science 2024 Anomaly detection in feature space for detecting changes in phytoplankton populations
21 papers in this thread — see all publications
Model debiasing and fairness
A network will happily learn the background instead of the object, and nothing in its accuracy on a held-out split will tell you it did.
We develop methods to discover biases nobody annotated in advance, and to remove them from models that have already learned them — including in medical imaging, where a shortcut learned from acquisition settings becomes a patient-safety problem.
- WACV 2027 Beyond the Generative Trilemma: Bias Amplification in Diffusion-GAN Hybrids
- CVPR 2026 Bias In, Bias Out? Finding Unbiased Subnetworks in Vanilla Models
- NeurIPS 2026 Diffusing debias: Synthetic bias amplification for model debiasing
- WACV 2026 How I Met Your Bias: Investigating Bias Amplification in Diffusion Models
- WACV 2026 Lose Your Self (LoYS): an adversarial entropy-based unsupervised approach for model debiasing
12 papers in this thread — see all publications
Learning from limited or distributed data
Labels are expensive, data often cannot leave the hospital that produced it, and the distribution shifts between where a model is trained and where it is used.
Self-supervised and transfer pre-training, few-shot adaptation, unsupervised domain adaptation and federated learning. This thread runs through the other three rather than sitting beside them.
- WACV 2027 PACER: Curriculum Self-Training with Spatial Uncertainty for Pseudo-Label Selection in Few-Shot Anatomical Landmark Detection
- MICCAI 2026 CDPM-Align: Multi-Scale Guidance-Aligned Diffusion Pretraining for Robust Few-Shot Anatomical Landmark Detection
- ICPR 2026 CutClean: Neural Network Pruning for Privacy-Preserving Inference
- ICML 2026 Distribution Alignment for One-Shot Federated Learning via Optimal Transport
- WACV 2026 Lose Your Self (LoYS): an adversarial entropy-based unsupervised approach for model debiasing
27 papers in this thread — see all publications
Funded projects
Next Waterfront
A deep learning model that assesses water quality from plankton imaging, camera video and sensor streams — reading the health of a body of water from the organisms living in it.
Elsewhere
Publications · Talks · Patents · Google Scholar · ORCID · GitHub