Biomedical engineer working on medical imaging, mostly brain MRI, in Python.
A lot of my projects start with a doubt about a good-looking result. Is the X-ray model actually looking at the lungs? Does the heart disease model still work at another hospital? How would you know a classifier got worse at a new site without labelling everything again? I'd rather find out than report the number and move on.
Research
- covid-xray-shortcut-learning: the classifier still scores 0.98 AUC with the lungs hidden. File metadata alone gets 0.99.
- breast-ultrasound-performance-monitoring: estimating sensitivity and PPV at a new hospital with few or no labels, on five datasets from five countries.
- Heart_disease_prediction: trained at one hospital, tested at three others. AUC drops from 0.90 to about 0.75, and the risk estimates are off by much more.
- ssl-brain-parcellation: self-supervised pretraining for whole-brain parcellation. It helps most when only 10 to 25 scans are labelled.
- Whole-Brain-Segmentation: a 3D U-Net that reproduces 44 FreeSurfer structures in seconds instead of hours.
- brain-age-prediction-ixi: brain age from T1 MRI, with age-bias correction and a look at scanner effects.
Teaching
- mri-recon-tutorial: MRI reconstruction from k-space to compressed sensing and deep learning, all in Colab.
- One brain slice through a CNN, a ViT, self-supervised and semi-supervised learning, with the actual numbers at every step.
- neural-networks-from-scratch and image-processing-from-scratch: the classic algorithms, written out in NumPy.
Small tools
segmentation-volumes · segmentation-colormap · csv-data-to-image
Python, PyTorch, MONAI, TensorFlow, scikit-learn, NumPy, nibabel