NeRF for Medical Imaging

This is a NeRF library that has been tested on medical images. Medical images have differences from the dataset that NeRF is typically trained on:

• No colour information
• No specular highlights
• Volumetric rendering of images follows different process

Installation

Install tiny-cuda-nn: https://github.com/NVlabs/tiny-cuda-nn Ensure that you have got the PyTorch extension working.

The code falls back to a standard PyTorch implementation, but is significantly slower without this extension.

Set up a conda environment and install the required packages:

conda create --name nerf --file conda-requirements.txt
conda install --name nerf --file conda-requirements.txt

Download walnut dataset from this webpage. The dataset is large, totalling around 4Gb. The tiff files should be stored in ./data/walnut/

Execution

conda activate nerf
python main.py [config_file]

Example config file in config/test_config.json Should take about 1 minute to train for 40 epochs results should be similar to /img/test_config

Training curves and slices will be output to out/{run_name}/

Config file

The JSON documentation lists configuration options.

TODO

• fix bug with jaw data
• add some .png files to img/ folder to show example output
• modify main.py to output info in out.csv