Synthetic data generation for modality-agnostic zero-shot promptable medical image segmentation.
Published by
SPIE Medical Imaging
Summary
Synthetic data generation for modality-agnostic zero-shot promptable medical image segmentation.
About
Highly accomplished AI Scientist specializing in medical imaging, leveraging expertise in self-supervised learning, foundation models, and synthetic data generation to drive innovation. Proven track record of developing advanced architectures and pipelines that significantly improve model performance, efficiency, and robustness for critical medical segmentation tasks. Eager to apply deep learning and imaging expertise to challenging problems in AI research and development.
Work
GE HealthCare
|Senior AI Scientist
San Ramon, CA, US
→
Summary
As a Senior AI Scientist, designed and developed cutting-edge synthetic data pipelines and 3D segmentation architectures for medical imaging, significantly advancing model performance and efficiency.
Highlights
Designed and implemented a synthetic data pipeline ~95% smaller than ImageNet, enabling entirely medical image-free training paradigms.
Developed a synthetic data generation framework that enabled foundation models to adapt without real medical data, outperforming zero-shot baselines like SAM and MedSAM on out-of-distribution datasets.
Outperformed baseline image quality enhancement models in 80% of evaluation cases by training on non-medical synthetic images, demonstrating anatomy-agnostic generalization across MR, CT, and X-ray.
Improved model convergence speed by 30% through pretraining self-supervised encoders with synthetic data, then fine-tuning for MR, CT, and X-ray segmentation.
Developed a 3D segmentation architecture that surpassed state-of-the-art (SOTA) performance, while being ~97% smaller and ~83% lower in FLOPS.
Developed an intensity-based data augmentation method, demonstrating wider coverage in latent space and improving robustness under distribution shift.
GE HealthCare
|AI/ML Ph.D. Intern
San Ramon, CA, US
→
Summary
As an AI/ML Ph.D. Intern, focused on finetuning advanced models and initiating synthetic data generation efforts to enhance medical segmentation tasks.
Highlights
Finetuned the Segment Anything Model for knee MRI segmentation, achieving SOTA expert model performance with a 0.85 Dice score.
Rapidly prototyped and comprehensively evaluated multiple approaches to exploit SAM's encoding and prompting capabilities, identifying optimal methods for performance.
Initiated the synthetic data generation effort at GEHC, enabling dataset-free finetuning and robust, generalized application for medical segmentation tasks.
Education
Columbia University in the City of New York
New York City, NY, United States of America
→
Ph.D.
Biomedical Engineering
Courses
Implemented a 100% speedup of the institutional brain imaging protocol and developed image denoising models, improving image quality (measured by signal-to-noise ratio) and accelerating Alzheimer's Disease imaging.
Trained Gibbs Ringing image artifact identification models in collaboration with radiologists for low-field and high-field brain MRI, achieving 0.878 ± 0.103 Cohen's kappa value with manual annotations, enhanced by Grad-CAM explainable AI for model trust.
Created PyPulseq (188 stars on GitHub), an open-source and vendor-neutral MR pulse sequence design framework widely adopted by multiple research labs for rapid MRI pulse sequence prototyping and democratization.
New York University
New York City, NY, United States of America
→
M.S.
Electrical Engineering
Publications
SynthFM: Training Modality-Agnostic Foundation Models for Medical Image Segmentation Without Real Medical Data.
Published by
IEEE International Symposium on Biomedical Imaging
Summary
SynthFM: Training Modality-Agnostic Foundation Models for Medical Image Segmentation Without Real Medical Data.
Accelerated MRI using intelligent protocolling and subject-specific denoising applied to Alzheimer's disease imaging.
Published by
Frontiers in Neuroscience
Summary
Accelerated MRI using intelligent protocolling and subject-specific denoising applied to Alzheimer's disease imaging.
Automated detection of motion artifacts in brain MR images using deep learning.
Published by
NMR in Biomedicine
Summary
Automated detection of motion artifacts in brain MR images using deep learning.
Developing and deploying deep learning models in brain magnetic resonance imaging: A review.
Published by
NMR in Biomedicine
Summary
Developing and deploying deep learning models in brain magnetic resonance imaging: A review.
Accelerated MRI using intelligent protocolling and subject-specific denoising.
Published by
ISMRM & ISMRT Annual Meeting & Exhibition
Summary
Accelerated MRI using intelligent protocolling and subject-specific denoising.
Feasibility of integrating a wearable accelerometer in very low-field MRI to detect motion.
Published by
ISMRM & ISMRT Annual Meeting & Exhibition
Summary
Feasibility of integrating a wearable accelerometer in very low-field MRI to detect motion.
Inline artifact identification using segmented acquisitions and deep learning.
Published by
Joint Annual Meeting ISMRM-ESMRMB & ISMRT 31st Annual Meeting
Summary
Inline artifact identification using segmented acquisitions and deep learning.
ArtifactID: Identifying artifacts in low-field MRI of the brain using deep learning.
Published by
Magnetic Resonance Imaging
Summary
ArtifactID: Identifying artifacts in low-field MRI of the brain using deep learning.
Intelligent denoising.
Published by
RSNA 2022 Scientific Assembly & Annual Meeting
Summary
Intelligent denoising.
A framework for validating open-source pulse sequences.
Published by
Magnetic Resonance Imaging
Summary
A framework for validating open-source pulse sequences.
ArtifactID: identifying artifacts in low field MRI using deep learning.
Published by
ISMRM & SMRT Annual Meeting & Exhibition
Summary
ArtifactID: identifying artifacts in low field MRI using deep learning.
PyPulseq in a web browser: a zero footprint tool for collaborative and vendor-neutral pulse sequence development.
Published by
ISMRM & SMRT Annual Meeting & Exhibition
Summary
PyPulseq in a web browser: a zero footprint tool for collaborative and vendor-neutral pulse sequence development.
Seq2prospa: translating PyPulseq for low-field imaging.
Published by
ISMRM & SMRT Annual Meeting & Exhibition
Summary
Seq2prospa: translating PyPulseq for low-field imaging.
Single board computer as a satellite-linked, deep learning capable pocket MR workstation: a feasibility study.
Published by
ISMRM & SMRT Annual Meeting & Exhibition
Summary
Single board computer as a satellite-linked, deep learning capable pocket MR workstation: a feasibility study.
Autonomous MRI.
Published by
Magnetic Resonance Imaging
Summary
Autonomous MRI.
PyPulseq: A Python Package for MRI Pulse Sequence Design.
Published by
Journal of Open Source Software
Summary
PyPulseq: A Python Package for MRI Pulse Sequence Design.
Virtual Scanner: MRI on a Browser.
Published by
Journal of Open Source Software
Summary
Virtual Scanner: MRI on a Browser.
Pulseq-Graphical Programming Interface: Open source visual environment for prototyping pulse sequences and integrated magnetic resonance imaging algorithm development.
Published by
Magnetic Resonance Imaging
Summary
Pulseq-Graphical Programming Interface: Open source visual environment for prototyping pulse sequences and integrated magnetic resonance imaging algorithm development.
Languages
English
Skills
AI
Self-supervised learning, Foundation models, Representation learning, Dataset-free learning.
Medical Imaging
MRI, CT, X-ray, Image reconstruction, Artifact correction.
Frameworks/Libraries
PyTorch, PyTorch Lightning, MONAI, NumPy, Pandas, Scipy, Scikit-learn, Nibabel.
Methods
Synthetic data generation, Data augmentation, Grad-CAM.