LoDoInd: Introducing A Benchmark Low-dose Industrial CT Dataset and Enhancing Denoising with 2.5D Deep Learning Techniques

Abstract: Computed Tomography (CT) is a widely employed non-destructive testing tool. In industrial applications, minimizing scanning time is crucial for efficient in-line inspection. One approach to achieve faster scanning is through low-dose CT. However, the reduction in radiation dose results in increased noise levels in the reconstructed CT images. Deep learning-based post-processing methods have shown promise in mitigating this noise, but their effectiveness relies on access to datasets with a substantial amount of training data. Existing low-dose CT datasets either are not specifically tailored for industrial applications or are based on simulated image formation. In this study, we present a new benchmark low-dose CT dataset, LoDoInd, which consists of experimental low-dose CT images explicitly designed for industrial purposes. LoDoInd incorporates complex and diverse secondary filling objects within the same testing object, simulating real-world scenarios encountered in industrial settings. The dataset can be accessed at https://zenodo.org/records/10356955. Building upon the foundation set by LoDoInd, we further investigate the efficacy of various post-processing methods in denoising tasks. Through a detailed comparative analysis of 2D, 2.5D, and 3D training, we demonstrate that 2.5D training strikes an optimal balance between performance and computational efficiency. This analysis showcases the potential of deep learning in improving the quality of low-dose CT images and also offers valuable insights for enhancing industrial CT applications with practical, efficient AI solutions. The corresponding code is available at https://github.com/jiayangshi/LoDoInd.