← All work

Case study 01 · 2026 · Harvard Medical School

Artery Segmentation & 3D Reconstruction

Comparing five deep learning models for carotid artery segmentation, and rebuilding coronary arteries in 3D from two X-ray views.

Year
2026
Field
Medical imaging
Stack
nnU-Net, SegMambaV2, PyTorch, Flask, React, Three.js
Result
0.92Dice score, carotid segmentation

Overview

As a Clinical AI R&D intern at Harvard Medical School (January to July 2026), I compared five deep learning models for carotid artery segmentation. nnU-Net and SegMambaV2 scored highest, with a Dice score of 0.92.

Dice measures how closely the predicted artery overlaps the expert-labelled one, from 0 (no overlap) to 1 (a perfect match).

I also built a pipeline that segments coronary arteries in X-ray angiograms and reconstructs them as 3D vessel models from two views. Validation checks measured mean absolute vessel-radius errors of 0.59 mm and 0.67 mm on two image pairs.

To make it usable, I built a Flask and React app where you upload DICOM scans, run the segmentation and explore the 3D vessels in the browser with Three.js.