LoDoPaB-CT, a benchmark dataset for low-dose computed tomography reconstruction.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning methods show promise in computed tomography (CT) image reconstruction.
• Comparing these deep learning approaches is difficult due to variations in training data and setups.
• A standardized, open-access dataset is needed for reproducible research and method comparison.

Purpose of the Study:

• To introduce the Low-Dose Parallel Beam (LoDoPaB)-CT dataset.
• To provide a comprehensive resource for training and evaluating deep learning and classical CT reconstruction algorithms.
• To facilitate advancements in low-dose CT imaging and reconstruction techniques.

Main Methods:

• Curated over 40,000 scan slices from approximately 800 patients from the LIDC/IDRI database.
• Simulated low photon count measurements for parallel beam CT.
• Detailed description of data selection and simulation setup, with publicly accessible generation scripts.
• Developed a Python library for simplified dataset access and an online reconstruction challenge.

Main Results:

• Established a large-scale, open-access dataset (LoDoPaB-CT) for CT image reconstruction research.
• The dataset supports the training and benchmarking of various reconstruction algorithms.
• Demonstrated the dataset's utility for transfer learning and sparse/limited-angle reconstruction scenarios.

Conclusions:

• The LoDoPaB-CT dataset provides a valuable resource for the medical imaging and AI communities.
• It enables standardized comparison of deep learning and classical reconstruction methods.
• Facilitates further development and validation of low-dose CT reconstruction techniques.