An exploration of compressed sensing fMRI time series with 3 different algorithms. Typically, compressed sensing reconstructs a single volume of MRI but fMRI are composed of many volumes; sensing along the time domain could reduce the number of volumes required. Of the 3 algorithms, BSBL-BO performed the best with the error curve elbowing around 30% subsampling.

Built a JetBot + an exploration and novice implementation of motion planning on said JetBot. This computational game theory project marked my first foray into optimization and a glimpse of its power muahahaha. It ain’t exquisite but it was heading in the right direction.

StoneAnno is my first published first-authorship paper, presenting at SPIE 2022. With the long-term goal of fully-automated robotic endoscopic surgery, we built a dataset of endoscopic kidney stone removal videos and investigated U-Net, U-Net++, and DenseNet for the segmentation task. We found a U-Net++ model that consistently achieves >0.9 Dice score, with low loss, and produces realistic, convincing segmentations. Moving forward, I am implementing our model on hardware for deployment in ORs, as a part of my master’s thesis, and I helped Dr. Kavoussi submit an R21 grant in October 2021.

I fell in love with dimensionality reduction when I was learning statistical ML. Since I also study neuroscience, I wanted to practice the art at the intersection of my interests. I compared the 3D projections of a 53-dimensional neurophysiology dataset produced by PCA and a shallow autoencoder.