A few projects that best represent where my work is right now, spanning sparse-view 3D reconstruction, gaze-aware wearables, and brain-computer interfaces.
Columbia University Irving Medical Center · under Prof. Yading Yuan
What & why
Investigated whether a patient's existing planning CT (already acquired and co-registered before radiotherapy) could serve as a better starting point than the standard noisy FDK reconstruction when using 3D Gaussian Splatting to reconstruct ultra-sparse-view CBCT: the kind of severely undersampled data produced when a single scan is split into breathing-phase bins for 4D motion-resolved radiotherapy imaging.
My contribution
Built the full experimental pipeline end-to-end: converting and normalizing the SynthRAD2025 CT/CBCT dataset, simulating sparse-view acquisitions via forward projection, designing a body-masked evaluation protocol to correct a background-air bias that was substantially overstating initial results, and independently validating the finding across two separate Gaussian Splatting codebases (R²-Gaussian and FaCT-GS) with paired statistical testing.
Outcome
Results were compiled into an abstract and supplemental manuscript submitted to SPIE Medical Imaging 2026 (Image Processing track). Decision is pending; it hasn't been accepted or published. Currently prototyping a follow-up extension (a NeRP-inspired two-stage training design that optimizes Gaussians directly against the planning CT before sparse-view fine-tuning) under active development with my PI.
MediaPipe · Depth Anything V2 · SAM
What & why
A video-passthrough prototype exploring gaze-contingent selective dimming: the idea that a wearable display could use real-time gaze tracking to detect what a person is actually attending to, then selectively dim or de-emphasize the rest of the visual field. Ties back to a broader interest in attention-aware wearables: if a system can tell where someone's focus is, it could reduce visual distraction in learning or work environments rather than presenting everything with equal visual weight.
My contribution
Building the core perception pipeline: iris/gaze tracking via MediaPipe, depth estimation using Depth Anything V2 (with MiDaS as a comparison/fallback), and segmentation via SAM to isolate the attended region from the rest of the scene. Also designing the multi-rate pipeline architecture needed to run these components together, since gaze tracking needs to update much faster than segmentation or depth estimation, so staging and synchronizing components at different rates is a core part of the engineering problem.
Current status
This is an early-stage prototype, not a finished system. The individual components (gaze tracking, depth estimation, segmentation) are working and are being integrated into the multi-rate pipeline, but there is no complete end-to-end demo, quantitative evaluation, or user testing yet. Best described as active, in-progress exploration rather than a completed project.
EEG-Based Speech Decoding with Semantic Comprehension Scoring · Neurohackathon, San Francisco, 2025
A non-invasive brain-computer interface that decodes what a person is reading from EEG brain signals and measures how well their brain is understanding each concept, aimed at adaptive, personalized education.
Headline results
66.7% top-10 word retrieval accuracy on EEG-to-text decoding (vs. 1.1% chance baseline), plus a trend toward the expected N400 semantic-violation effect on real EEG data (p = 0.097, not significant at conventional thresholds).
Read the full case study →