CTML

This is a python package for ML approaches for handling Micro-CT Data. Functionality will be expanded over time.

Current functionality is limited to a denoising approach for reconstructed data using an unsupervised (inference-only) model.

Inference Model Training

Procedure:

1. First, an Ansecombe transform is applied to the noisy images with correlated noise (G+P).
2. Since the inter-slice width is negligible, the change in signal between two consecutive slices is approximately 0.
3. However, noise varies between the two slices. Patches from slice[i] were used as input and the corresponding slice[i+1] patch was used as target.
4. Over iterated epochs, the unet learns the mean of the observations (true signal) and produces a clean image.

Two pretrained weight sets are available:
module/denoise_phase.pt
module/denoise_nophase.pt
Both samples were trained on a dataset of 4976 13568x13568 reconstruction slices. The first was trained on a dataset with phase retrieval[1] applied, the second was not.

Usage:

python ctml --data-dir data/train --weights data/final.pt --cuda utraining --nb-epochs 100

For full paramter listing see
python ctml --help
python ctml utraining --help

Model Application

The finalized model is applied to the input file in patches, that are merged together.

Usage:

python ctml --data-dir data/train --weights data/denoise_phase.pt --cuda udenoise --output-dir=data/clean

For full paramter listing see
python ctml --help
python ctml udenoise --help

(Note that the batch-size parameter listed in python ctml --help is ignored here due to limitations of empatches).

Current Limits:

1. Data is assumed to be float32.
2. Model cannot remove ring artifacts.
3. While CUDA is technically not required, model is too computationally intensive for straight CPU usage.

Installation:

1. Install CUDA as appropriate for your OS, if not already installed.

2. Create and activate a conda environment, which is necessary for use of CUDA GPU tools.

3. Install Pytorch. Links below as per time of writing.

   1. CUDA >= 11.7: Current Version
   2. CUDA < 11.7: Old Version

   As this will limit available python versions and is hardware dependent, should be done before installing other requirements.

4. Install requirements via conda or pip:
   conda env update -f ctml.yaml
   pip install -r requirements.txt

   Note that pytorch is required but not listed in the requirements due to variance in CUDA version requirements.

[1] Paganin, David, et al. "Simultaneous phase and amplitude extraction from a single defocused image of a homogeneous object." Journal of microscopy 206.1 (2002): 33-40.