Yiheng Li

• Deep learning application in medical imaging
• Multimodal data fusion
• Image registration and quality control
• Self-supervised learning
• Open source software development
• LLM, LoRA, Stable diffusion and AIGC in general

EXPERIENCES

Deep Learning Research Scientist at Subtle Medical Inc.

Menlo Park, CA 08/2021-07/2023

Deep Learning based Image Co-Registration Quality Control

08/2021-02/2022

• Built a self-supervised classification model to QC the co-registration performance for pairs of 3D MRI images.
• Developed a data augmentation pipeline including affine and deformable movement to generate pseudo data pairs to tackle the data challenges of obtaining labelled well-registered MRI data pairs.
• To achieve multi-contrast prediction, compared two technical approaches: augmented contrasts vs. modality invariant representation.

Meeting accepted: ISMRM 2022, ASNR 2022, RSNA 2022

Deep Learning-Based Multi Contrast MRI Registration Model with a Realistic Flow Field and Reduced Over-Smoothing Effect

02/2022-09/2022

• Developed a deep learning based co-registration model that can be applied to multi-contrast MRI images of multiple anatomies.
• Novel attempts, using “Jacobin loss” and “cycle consistent loss”, to deal with unrealistic flow field and over-smoothing effect of deep learning- based registration methods, especially in the “VoxelMorph” and “SynthMorph” framework.
• Converted TensorFlow based “SynthMorph” code and trained PyTorch versions of “SynthMorph” and other variations of the model.
• Improved the SSIM and PSNR of the registered image on BraTS and Lumbar-Spine Open Dataset by ~40% and ~50% respectively.

Meeting accepted: ISMRM 2023

A Self-Supervised Key Point Detection Framework For Multiple Applications

09/2022-07/2023

• Explore, adapted, developed and assessed multiple technical workflows for self-supervised key point detection in medical images.
• Developed a two-step rule-based pipeline for brain MRI auto- formatting, using ANTs affine registration and SIFT key point matching.
• Guiding a intern to adopted and optimized the performance of the “KeyMorph”, an automatic key point generator with registration training mechanism, by customizing the loss with distance and applying additional mask.
• Start with the replication of a real-time self-supervised key point detector paper in PyTorch Lightning. Optimize over the original paper’s result by ~30% on a private self-curated test dataset with modified training strategy and grid loss search. A multi-purpose training framework is introduced by adding a tail to provide binary prediction of the key point existence. Transformer-based encoder replaced the CNN based encoder. Reinforced learning and iterative predictions are both tested for finer prediction.

Creation and Maintenance of a PyTorch-Lightning and MONAI based Deep Learning Training, Logging and Inference Helper Package: Lumos-ToolKit

09/2022-07/2023

• Created and maintained a toolkit which includes pipelines for the key steps and pain point in deep learning model development for medical imaging: dataset curation and testing; image affine, deformable and other spatial transformations; logging and documentation of the model hyper-parameters, figures and performances; management of the training settings and configurations; image preprocessing pipeline; complicated loss settings and combinations of losses.
• The whole package dynamically integrates the following packages: “pydantic”, “argdantic”, “PyTorch-Lightning”, “MONAI”, “rich”, etc.

Education

Stanford University

09/2019-06/2021

M.Sc. in Biomedical Informatics

Shanghai Jiao Tong University

09/2015-06/2019

B.Sc. in Resource and Environmental Science

University of California, Berkeley

01/2018-05/2018

International Exchange Program