[Paper] [Dataset] [BibTeX]

The XctDiff is capable of reconstructing CT images with consistent anatomical structures from a single radiographic projection image. This will be able to be extended to more meaningful work, such as quantitative body composition analysis, expanding medical datasets, and so on.

conda env create -f environment.yml

First, we need to train a 3D perceptual compression encoder:

python main.py --cfg_path models/autoencoder/autoencoder_vq_32x32x32_8.yaml --gpus=2 --max_steps 80000

Then, we need to train an encoder to convert X-ray images to 3D features:

python main.py --cfg_path models/embedding/embedding_32x32x32_8.yaml --gpus=2 --max_steps 50000

Finally, we integrate the two components and jointly train a latent generative model

python main.py --cfg_path /home/first/XctDiff/models/ldm/ldm_32x32x32_8.yaml --gpus=2 --max_steps 100000

It is worth noting that for training, we need to specify two component pre-training weight files in the configuration file ckpt_path

python .py --cfg_path models/ldm/ldm_32x32x32_8.yaml --input_path demo/1002.png  --ckpt_path checkpoint/xctdiff.ckpt