About DL-BeamHardening

This repository contains the work related to the deep-learning based beam hardening removal.

Setup-Project:

Install Python packages:
pip install torch pytorch-lightning torchmetrics matplotlib h5py numpy einops pyqt5

Build Voxie & Model-load extension:
tools/build.sh

Repository-Structure:

📦dl_beamhardening
┣ 📂aRTist_project
┃ ┣ Example aRTist projects with IPVS settings
┃ ┗ 📂Spektren
┃ ┃ ┗ Mono & poly (IPVS-Setup) spectrum files
┣ 📂data_analysis_scripts
┃ ┗ Scripts for calculating mean gray value, data analysis & .tiff plot
┣ 📂dataset_preparation_scripts
┃ ┣ 📂preprocessing
┃ ┃ ┗ Script for flatfield correction
┃ ┣ 📂tiff_to_hdf5
┃ ┃ ┣ 📂stuff
┃ ┃ ┃ ┗ Various debug, plot & volume compare scripts
┃ ┃ ┗ Tiff to hdf5 converter script
┃ ┗ Script for transposing & cutting volumes
┣ 📂doc ┃ ┗ OneNote notebook & exported OneNote pages
┣ 📂models
┃ ┣ 📂checkpoints
┃ ┃ ┗ CNN-AI-CT checkpoints/traces that can be loaded by voxie
┃ ┣ 📂test_cases
┃ ┃ ┗ unittest, to check dataloader logic
┃ ┣ 📜CNN_ai_ct.py
┃ ┣ 📜CNN_ai_ct_silu.py (CNN-AI-CT with SILU-Activations)
┃ ┣ 📜CNN_ai_ct_skip.py (CNN-AI-CT with SKIP-Connections)
┃ ┣ 📜CNN_ai_ct_trans_skip.py (CNN-AI-CT with SKIP-Connection & extra Linear-Layers)
┃ ┣ 📜IRR_CNN_ai_ct.py (CNN-AI-CT with Iterative-Residual-Refinement)
┃ ┣ 📜PerceiverModel.py (DeepMinds Perceiver-Model)
┃ ┣ 📜Unet.py
┃ ┣ 📜dataloader.py
┃ ┣ 📜datasets.json (Contains path of all datasets)
┃ ┣ 📜debug_datasets.json
┃ ┣ 📜noisy_indexes.json (Contains indices of slices that contain only noise 256x256 patches)
┃ ┣ 📜parseTorchScript.py
┃ ┣ 📜removeBeamHardening.py (Remove Beam-Hardening from volume with passed checkpoint)
┃ ┣ 📜runModel.py (Script to load & test checkpoints)
┃ ┣ 📜train.py (Main script to start training & choose models)
┃ ┣ 📜utils.py
┃ ┗ 📜visualization.py
┣ 📂src
┃ ┗ Extern voxie filter to load pyTorch models (traces) of CNN-AI-CT
┣ 📂subprojects
┃ ┗ 📂voxie-intern
┣ 📂test_data
┃ ┗ small volume data for test-purposes
┣ 📂tools
┃ ┗ build script (Voxie + extension) & normalisation script
┣ 📜.clang-format
┣ 📜.gitignore
┣ 📜.gitmodules
┣ 📜README.md
┗ 📜meson.build

Use-cases naming convention

To distinguish different training runs & model checkpoints

Training & RunModel

For training and checkpoint loading details/ scripts check this folder on pasnas:

/net/pasnas01/pool1/enpro-2021-voxie/training/slurm_files

This folder contains all used slurm and corresponding bash scripts.

Dataset preparation scripts

tiff_to_hdf5

Converts a bunch of Tiff files into a single HDF5 file and appends meta-data needed for the reconstruction. >Manual flat-field correction can be applied if needed.

Run: (Without flat-field correction):

<python-interpreter> tiff_to_hdf5.py -f <path-tiff-file-folder> -o <absolute-path-of-output> -dsra <distance-source-rotation>

Run: (With flat-field correction):

<python-interpreter> tiff_to_hdf5.py -f <path-tiff-file-folder> -o <absolute-path-of-output> -dsra <distance-source-rotation> -w <path-to-white-tiff-image>

Hint:

• distance-source-rotation must be in meter [m]
• white-image: Ct-projection image without an object and detector noise but apart from this with the same parameters as the normal projections

reco_post_processing.py

Transposes and cuts the monochromatic and polychromatic >reconstructed volumes identical.

Run:

<python-interpreter> reco_post_processing.py -fp <path-reconstructed-volume-poly-hdf5> -fm <path-reconstructed-volume-mono-hdf5> -op <output-path-poly-hdf5> -om <output-path-mono-hdf5> -fc <cutting-coefficient>  

Hint:

• cutting-coefficient: The factor is multiplied with the mean_grey_value of the volume. Slices are then removed if the mean slice gray value is lower as this product.

Dataset analysis scripts

plot_sample_differences.py

Create images of the slices that can be used to analyze the data: Poly, Mono, Poly-Mono & Histogram(Poly-Mono)

Run:

<python-interpreter> plot_sample_differences.py -f <dataset-json-file-path> -o <output-path> -ss <step-size> -dn <names-of-datasets>

Hint:

• step-size: Each int(dataset_size/step_size) sample is written out
• names-of-datasets: Default is all
• output-looks like:

calculate_cutted_mean.py

Calculates the mean gray value of a volume and appends it as meta-data.

calculate_cutted_mean.py -f <dataset-json-file-path> -dn <names-of-datasets>

Hint:

• names-of-datasets: Default is all