JEB1433

Hip Arthroscopy Image Registration This repo is what was used for the JEB1433 project performing 2D/3D image registration on a hip arthroscopy dataset from Johns Hopkins University

There are 4 notebooks labelled as follows:

1. Setup - downloads required repositories and load dataset
2. DRR Generation - Check that diffDRR is functioning
3. Regsitration - there are 2 versions one regular for one seed and another that loops and trials translations and rotations individually
4. Animation - prepares .gif that was used for the JEB1433 presentation

The final python notebook was the one used in google colabs to complete the iterations on every patient's dataset.

The necessary background data and functions are detailed below:

Dataset

The pipeline uses DeepFluoro (Grupp et al., IPCAI 2020):

• 6 cadaveric pelvis specimens
• Paired CT volumes + real C-arm fluoroscopy images
• Ground truth camera poses (from offline registration)
• Dataset auto-downloads on first run via diffdrrdata (~165 MB, 4× downsampled)

Full dataset: https://doi.org/10.7281/T1/IFSXNV

---

The DRR Stack

We use DiffDRR (Gopalakrishnan & Golland, 2022) which reformulates Siddon's ray casting algorithm as vectorized PyTorch tensor operations:

• GPU-accelerated: Sub-second renders at 512×512 (vs 27–52 seconds in Freund 2004)
• Auto-differentiable: Gradient flows from image similarity → DRR pixels → camera pose
• Siddon's method: Exact radiologic path integral (not interpolated like projection fields)

This enables gradient-based optimization rather than derivative-free Powell's method.

---