I’m a postdoc at MIT CSAIL in Polina Golland’s Medical Vision Group where I work on representation learning frameworks for zero-shot generalization to entirely new tasks, biomedical contexts, and datasets. I want to build machine learning methods that biomedical domain experts can use as-is without needing to retrain or adapt to their use-cases.
I got my Ph.D. from NYU under Guido Gerig at the VIDA Center where I worked on generative models and inverse problems in medical image analysis contexts.
I’m always happy to talk, feel free to reach out if our interests overlap!
Research Overview
I tackle the pervasive lack of biomedical training data and supervision to create models that generalize to new biomedical settings with minimal to no new training data.
This has included:
- Learning representations from randomized image synthesizers, enabling zero-shot generalization to any biomedical modality and organism without any domain-speicific training data.
- Developing equivariant layers for sparse spatio-spherical deconvolution in diffusion MRI and creating data-efficient GANs for digital pathology with limited annotated subjects.
- Introducing dense multi-scale representation learning methods for neuroimaging tasks such as multi-modality non-rigid registration and one-shot segmentation in settings with ample unannotated training images.
Recent Updates
10/24 | New work! Using random compositions of templates to drive a self-supervised objective, we get entirely general-purpose representations that yield SOTA 3D registration and segmentation. |
09/24 | New at NeurIPS 2024: fast equivariant modeling of spatio-spherical data leads to SOTA neuronal fiber tracking in diffusion MRI. |
07/24 | I’ll be an area chair for ML4H 2024. |
06/24 | Our new work (led by Benjamin) in IEEE TMI shows that intensity-canonicalization greatly improves SE(3)-equivariant rigid tracking in volumetric data. |
02/24 | At CVPR 2024: Our new work (led by Vivek) leverages differentiable rendering, se(3) lie algebras, self-supervised geodesic pre-training and more to get sub-millimeter accurate 2D/3D rigid registration, all in about a second. |
12/23 | I’m a guest on this week’s AI-ready Healthcare podcast! Listen here. |
12/23 | I’ll be an area chair for CHIL 2024. |
09/23 | I was an outstanding reviewer for ICCV 2023 and an honorable mention for outstanding reviewer for MICCAI 2023. |
07/23 | Our work on training a single model to segment 3D blob-like instances without retraining in any 3D bio-microscopy or radiology dataset is now out and accepted by WACV 2024! Try it in Colab. |
07/23 | Our MIDL oral papers were both selected for journal extensions in Medical Image Analysis and Nalini won best oral paper! |
05/23 | Our work (led by Neerav) on better calibrating neural nets for spatial tasks was early accepted to MICCAI’23! Preprint here. |
05/23 | I was an outstanding reviewer for CVPR 2023. |
03/23 | We have two papers accepted as oral talks at MIDL 2023! We build E(3) x SO(3) equivariant networks for diffusion MRI and develop data-consistent MRI motion correction. |
02/23 | Our work on learning probabilistic piecewise rigid atlases for model organisms used in neuroscience was accepted as an oral talk by IPMI 2023! Preprint here and code here. |
10/22 | I was an outstanding reviewer for ECCV 2022. |
10/22 | I gave a series of talks on multi-scale dense representation learning at MIT CSAIL, the Martinos Center, and Brigham and Womens Hospital. |
09/22 | Our work on developing multi-scale and locality-sensitive self-supervised representation learning stategies was accepted as an oral talk by NeurIPS 2022! |
08/22 | Our work building transformers for MRI reconstruction was accepted by WACV 2023. |
07/22 | Our work on unsupervised MRI reconstruction was accepted by Medical Image Analysis. |
06/22 | I received a student travel award for MICCAI 2022. |
06/22 | Our work on representation learning for multi-modality non-rigid registration was accepted to MICCAI 2022. Preprint here. |
05/22 | I was an outstanding reviewer for CVPR 2022. |
04/22 | I received the Pearl Brownstein Doctoral Research Award from NYU CSE for “doctoral research which shows the greatest promise”. |
07/21 | Our paper on learning deformable templates was accepted by ICCV 2021! Head over to our webpage for the preprint, code, and highlights. |