Low-dose 3D cone-beam CT (CBCT) reconstruction challenge

About

• Team member names:
  • Sasidhar Alavala (MS(R), IIT Tirupati, India)
  • Dr. Subrahmanyam Gorthi (Asst. Professor, IIT Tirupati, India)
• Abstract:
  • Our approach integrates SwinIR-based sinogram enhancement module, coupled with Nesterov Accelerated Gradient Descent (NAG) for solving the least squares problem in CT image reconstruction. To address the challenge of excessive blurring during reconstruction, a second phase of image processing is done. This involves using another SwinIR-based CT enhancement module for enhancing features that may have been compromised in the reconstruction process. The combination of sinogram enhancement and CT enhancement modules aims to provide a better solution for low-dose and clinical-dose CBCT reconstruction, offering improved image clarity and fine detail preservation.

Model Zoo

Please go to MODEL HUB for model weights.

Usage

Install

You can use either the conda environment file to install dependencies or create a new environment and install the mentioned packages:

• Creating conda test enivronment with YML file:

conda env create -f conda_environment.yml

• Creating conda test enivronment without YML file:

conda create -n test_env python==3.10.12
conda activate test_env

Then install torch==2.1.0+cu121 torchvision==0.16.0+cu121 cudatoolkit=11.3.1 numpy==1.26.0 timm==0.9.12 astra-toolbox=2.1.2 tomosipo==0.6.0 ts-algorithms==0.1.0

Data preparation

The data and model weights folder structure is as follows:

$ tree data
data
├── sino_test_low
│   ├── 0901_sino_low_dose.npy
│   ├── 0902_sino_low_dose.npy   
│   └── ...
├── sino_test_clinical
│   ├── 0901_sino_clinical_dose.npy
│   ├── 0902_sino_clinical_dose.npy   
│   └── ...
├── ct_groundtruth
│   ├── 0901_clean_fdk_256.npy
│   ├── 0902_clean_fdk_256.npy   
│   └── ...
├── ct_output_low
├── ct_output_clinical
└── model_zoo
    ├── low_sino_231.pth
    ├── low_ct_117.pth
    ├── clinical_sino_148.pth   
    └── clinical_ct_186.pth

Evaluation

To evaluate on the test dataset run:

python3 test_low.py
python3 test_clinical.py

References

• Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., & Timofte, R. (2021). Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 1833-1844).
• Hendriksen, A. A., Schut, D., Palenstijn, W. J., Viganó, N., Kim, J., Pelt, D. M., ... & Batenburg, K. J. (2021). Tomosipo: fast, flexible, and convenient 3D tomography for complex scanning geometries in Python. Optics Express, 29(24), 40494-40513.
• Nesterov, Y. (2003). Introductory lectures on convex optimization: A basic course (Vol. 87). Springer Science & Business Media.

License and Acknowledgement

This work is made public under the MIT license. The codes are based on SwinIR and ts_algorithms. Please also follow their licenses. Thanks for their awesome works.