This repo hosts the app we developed during the 2025 Rakathon.
The basic idea is to allow accurate registration and sampling between large reference CT scans and smaller but higher-resolution treatment-time scans recovered from the conical beam scanner.
How?
- Interpret each CT scan as a 3D signal with given intensities
- Use 3-Dimensional FFT (see file
pointcloud_alignment/fourier.py) to estimate the 3D translation vector between these pointclouds - Align them into same-sized, equally-spaced grid volumes
- Perform local SGD-based optimization to align complex shapes inside of the CT scan structures.
A crude visualizer written in pure JS is available.
If you want to try it out, you can do:
uv pip install . && python app.py --sample SAMPLE_001
Assuming you have the dataset. The dataset can be preprocessed using the script
dicom_processing/organize_ct_by_date_frame.py.
Some screenshots:
