Overview of all projects by Michael Kim (kimm58)
Github: https://github.com/MASILab/Diff_MRI_Harmonization.git
Data: backup via amazon (ADSP)
This was my first research project, looking at performance of the ComBat algorithm for site-wise harmonization when using Diffusion MRI data. We assessed performance using a well-matched dataset from two different sites (VMAP, BLSA) as a baseline, then performed a bootstap analysis of how varying sample size, sample size imbalance, and covariate shift/overlap affect the estimates of association of Age with fractional anisotropy (FA) and mean diffusivity (MD), and the site-estimated shift and scale of the harmonized features (which were the FA/MD of regions in the EVE3 WMAtlas). Our analyses demonstated that ComBat performance is worse than expected for diffusion MRI data (i.e. need a larger sample size and tighter covariate overlap) in order to get estimates from ComBat that do not deviate too far from the well-matched scenario.
Github: https://github.com/MASILab/DeformableSurfaceReg.git
Data: XNAT: NeRF_Surface_Landman: https://xnat.vanderbilt.edu/xnat/data/projects/NeRF_Surface_Landman
Paper: https://medicalimaging.spiedigitallibrary.org/proceedings/Download?urlId=10.1117%2F12.3005578&downloadType=proceedings%20article&isResultClick=True https://www.biorxiv.org/content/10.1101/2023.09.22.558989v1
SPIE 2024 paper - A characterization of low-cost photogrammetry techniques for surface reconstruction. Using surfaces obtained from CT images as the ground truth, we assess the quality of the photogrammetry surfaces obtained from Neural Radiance Fields (NeRF, as instant-ngp implementation) and the VECTRA H1 3D Imaging System of a hip implant, a knee implant, and a steak. First, we register the photogrammetry surfaces to the ground truth CT. We then use fiducial registration error and average surface distance as metrics to assess quality. We find that the H1 camera provides smooth surfaces, but cannot reconstruct objects that are larger or have sharp changes in geometry. Whereas NeRF has the capability to capture finer detail, the surfaces are much noisier and more artifact-prone. Our conclusion was that NeRF-based methods have the greater potential, but improvements need to be made to the implementation.
3.) Characterization of Neural Radiance Field-Based 3D Reconstruction of Organ Tissue Surfaces (technical paper)
Github: https://github.com/MASILab/DeformableSurfaceReg.git
Data: XNAT: NeRF_Surface_Landman: https://xnat.vanderbilt.edu/xnat/data/projects/NeRF_Surface_Landman
Extension of SPIE 2024 - Characterization of NeRF-based techniques for reconstruction of human organ tissue surfaces. We assess multiple points in the processing pipeline (image acquisition and surface reconstruction) to determine what aspects of organ surfaces make reconstruction difficult and what solutions work best to optimize surface reconvery. We find that addition of external features in the image scene help with feature matching of the images, particularly when only adding a few eternal fiducials. We also find that modification of the specular estimator of the nerf2mesh implementation did not provide better surface recovery. However, having a high laplacian regularization helps remove/alleviate areas of low-density estimation. Building a shape model to fix the estimated density field also did not provide a successful solution.
Github: https://github.com/MASILab/Informatics_ADSP.git
Data: Used MASIVar (timing data is in repo as well)
SPIE 2025 - This paper is an overview paper for the ADSP processing methodology, descibing how we designed our data storage and processing to fit within specified design criteria of large-scale neuroimaging data maintenance. The criteria include (but are not limited to) efficiency, cost-effectiveness, and reproducibility. We demonstrate how our method performs similarly to other options while maintaining a nearly 20x lower cost-performance ratio, while remaining low complexity on our end for manual effort.
Github: https://github.com/MASILab/Diff_MRI_Harmonization.git (and R01_Data_Organization)
Organized and preprocessed the majority of the datasets for the ADSP/R01 grants. Created a script generator that allows for easy and standardized preprocessing of data on ACCRE (or locally as well), given that data is in BIDS format. Also created a few query tools that do various tasks, such as summarize the preprocessing that has been run, number of shells, etc.
Github: https://github.com/MASILab/ADSP_AutoQA.git
Code for automated QA of processing pipelines in the ADSP/R01 project. There are 4 ideas/problems that this tool is trying to address: 1.) Promote consistency of data QA, so that everyone on the team is performing QA in the same way. 2.) Facilitates aggregation of QA results across pipelines and datasets, as changes are automatically pushed to a CSV file that indicates the QA status. Not only does this reduce the manual effort in keeping track of the QA status of processing, but makes it easier to share QA results within lab and with collaborators who use the data. 3.) Gives us a structured organization for the QA of our pipelines/datasets, so that we can easily combine the QA results of multiple pipelines/datasets together. Finally, 4.) it provides a nice graphical user interface for reporting the QA status.
Github: https://github.com/MASILab/AtlasToDiffusionReg
Created a singularity for deformably registering an atlas to a diffusion scan of a subject using ANTs, fsl, c3d, etc. Requires a T1, a segmentation of the T1 (SLANT preferrably), the preprocessed DWI scan, the structural template and the labelmap/atals, and label files for both the segmentation and the labelmap/atlas. Outputs the transformations and registered atlas, as well as a CSV of the mean, median, and stdev of the tensor metrics for FA, MD, AD, RD.