Research
I work in the Ferradal Lab at Indiana University on diffuse optical tomography (DOT), which images brain hemodynamics (HbO and HbR) from near-infrared light measured on the scalp. DOT is portable and safe at the bedside, which makes it a strong fit for the neonatal brain. Its images are only as good as the model of light travelling through the head, so my work is about making that model accurate for each subject and fast enough to use. The piglet is our model of the neonatal brain, in a collaboration with the University of Pennsylvania and Children's Hospital of Philadelphia on cerebral monitoring around cardiac arrest.
Does the head model matter?
I built an end-to-end pipeline from a subject's T1 MRI to a DOT reconstruction: brain extraction and tissue segmentation (extracerebral tissue, CSF, gray and white matter) with nnU-Net and SynthSeg-style synthetic training, a finite-element head mesh with iso2mesh, then forward models and wavelength-specific Jacobians (sensitivity matrices). The sensitivity maps quantify cortical coverage and guided the design of a 57-optode high-density array. The question this thread asks is how much a subject-specific head model changes the reconstructed image compared with an atlas.
I presented the system as a first-author oral at the Optica Biophotonics Congress 2026: "Developing Diffuse Optical Tomography for Peri-Arrest Monitoring in Piglets" (OT2D.3). A companion poster with Brian White's group, "First-in-Large-Animal Diffuse Optical Tomography: Hardware and Pipeline for Peri-Arrest Monitoring in Piglets," is at fNIRS 2026 in Macau in October 2026.
How fast can a subject-specific forward model be?
A head model per subject is only practical if the forward model is cheap. I moved the DOT forward solve to the GPU with CUDA and cut one step from about 1400 s to 8 s, validated against the MATLAB reference. I work with both Monte Carlo photon transport (MCX) and finite-element solvers (NIRFAST/NIRFASTer, NeuroDOT).
Can one brain-extraction model work across species?
Every head model starts with finding the brain in the MRI. Tools built for adult human scans often fail on piglet and other animal MRI, which come in many contrasts and shapes. I am training segmenters on synthetic, domain-randomized images, aiming for one model that handles new scans across species.
Code
Code built on unpublished lab data stays private until the work is published. My public code and open-source contributions are on the Projects page.