Question regarding irregularly spaced angles and small number of images

[jqmcginnis]

Hello 🙂

Congratulations on the NeurIPS paper, and thank you very much for providing this excellent codebase.

Looking at the Readmet, you mentioned that we can also use our data to reconstruct the CT volumes. I have a unique dataset, with only four projections from non-symmetrically (i.e. irregularly) spaced angles (they do not even occupy 360 degrees). Given that you state that you would only recommend this for >100 x-ray images, I wanted to ask if you think your framework supports such a small number of images and if it can also handle non-uniformly spaced projections?

[Ruyi-Zha]

Hi, thanks for your interest in our work!

Yes, our code is compatible with an arbitrary number of projections and irregularly spaced angles. You can adjust these parameters by editing the information in the meta_data.json file.

"proj_train": [
        {
            "file_path": "YOUR_DATA.npy",
            "angle": YOUR_ANGLE
        },
       ...
]

The recommendation of using >100 views is primarily for generating a pseudo-ground truth volume for quantitative evaluation. In our real-world experiments, we subsample to 25, 50, or 75 views for training. You can use any number of projections, though our method may not perform well in extreme spase-view cases (<25 views), because it is self-supervised without pretrained deep learning priors. However, since our method is differentiable, the performance could potentially improve by integrating additional networks. While this direction is beyond the scope of this work, we encourage you to explore it.