BoneEnhance

Methods for improving image quality on computed tomography -imaged bone

(c) Santeri Rytky, University of Oulu, 2021-2023

Background

Clinical cone-beam computed tomography (CBCT) devices are limited to imaging tissues of submillimeter scale. This repository is used to create super-resolution models trained on high-resolution micro-computed tomography (µCT) images. For a detailed description of the method, refer to the publication by Rytky SJO et al.

Prerequisites

• Anaconda installation

git clone https://github.com/MIPT-Oulu/BoneEnhance.git
cd BoneEnhance
conda env create -f environment.yml

Usage

Model training

• Create a training dataset: Use the script create_training_data.py to simulate image pairs from high-resolution µCT scans. Set the data and save paths as well as resolution and save parameters at the beginning of the script.

• Set the path name for training data in session.py (init_experiment() function)

• Create a configuration file to the experiments/run folder. Example experiments are included in the folder. All experiments are conducted subsequently during training.

conda activate bone-enhance-env
python scripts/train.py

Inference

For 2D prediction, use inference_tiles_2d.py. For 3D data, use inference_tiles_3d.py. Running inference_tiles_large_3d.py allows to run inference on larger samples, and merge on CPU. Using inference_tiles_large_pseudo3d.py allows merging 2D predictions on orthogonal planes. Update the snap variable, image path and save directory.

License

This software is distributed under the MIT License.