Development of a modular and open-source tomographic imaging software: enhancing the reconstruction module for low-dose ct and dbt

Abstract

This study aims to develop an open source, modular, system simulation software that can be used in both two and three dimensions. The focus of this software is the recoding sub-module in tomographic image reconstruction. Tomographic image reconstruction is a critical process that aims to create three-dimensional images using raw projection data. Medical imaging techniques such as Computed Tomography (CT) and Digital Breast Tomosynthesis (DBT) play an important role in the early diagnosis and treatment of patients. However, these techniques have many limitations. The X-ray radiation used can cause DNA damage at the cellular level and increase the risk of cancer. Especially for patients who undergo frequent screening and pediatric patients, the importance of low-dose screening is even more crutial. Working with low-dose data brings with it difficulties such as decrease in image quality and increase in noise level due to the fact that projection data is obtained with few photons. Therefore, advanced reconstruction algorithms are needed to obtain images of good quality.When the existing system simulations were examined, three main tool sets were seen to stand out: ASTRA, TIGRE and LAVI. Each of these toolboxes is focused on specific areas. ASTRA is optimized specifically for small and medium-sized problems with a combination of Python and C++. It is widely used in tomographic imaging experiments and offers important support to advanced algorithms. However, it requires limited scalability for very large data sets and manual optimization for large-scale problems. The user interface is complex and not suitable for adding different modules. TIGRE is designed on MATLAB and Python platforms, hence it appeals to a wider audience in the field of computed tomography. It stands out with multi-GPU usage and advanced memory management. However, it has disadvantages such as complexity in multiple GPU setups and dependence on CPU RAM in large data sets. Also, it is not modular. LAVI is used exclusively for digital breast tomosynthesis and is installed in MATLAB. It uses advanced memory management techniques with CUDA in the processing of large data sets. It offers user-friendly visualization tools. However, the software is not flexible in terms of design, and performance improvements are still ongoing. This study goes beyond all these toolboxes and develops a modular system simulation software that offers a wide range of solutions for both CT and DBT. This toolbox, which is open source and can be used by different academic teams in their studies, adapts to different geometric scenarios with different coordinate system options and various types of sources and detectors. In addition, it offers a design that is suitable for cross-use in the reconstruction process by accommodating various forward and back-projection algorithms. It aims to minimize noise with analytical and statistical reconstruction algorithms and regulators. This study, which emphasizes the hardware effect in software design, also provides a simulation environment suitable for artificial intelligence-based back-projection algorithms. A key aspect of the software is its comprehensiveness in the Fourier domain. The software addresses challenges such as limited angular ranges and the incomplete sampling of the Fourier domain, which can lead to a decline in image quality. By supporting a wide range of geometric configurations and enabling precise Fourier domain sampling, the software allows for higher-quality reconstructions, even in complex imaging scenarios. The software utilizes the principles of the Fourier Slice Theorem to explain how projection data is sampled in the frequency domain, facilitating more accurate and detailed image reconstructions. The developed modular system simulation software is designed with flexibility, inclusivity, extensibility, and user-friendliness in mind. Performance optimizations have been achieved by parallelizing Distance-Driven (DD) and Branchless Distance-Driven (BDD) algorithms, reducing the initial processing time from 32.2 seconds to 7.17 seconds with CPU parallelization, and further to 0.008 seconds with GPU parallelization. These results underscore the critical role of parallel processing in enhancing the efficiency of image reconstruction, particularly when working with large data sets. Evaluation metrics such as Structural Similarity Index (SSIM), Root Mean Square Error (RMSE), and Peak Signal-to-Noise Ratio (PSNR) were used to assess the quality of reconstructions, demonstrating the software's capability to produce high-accuracy reconstructions in both analytical and clinical settings. The software's design also emphasizes modularity and object-oriented programming, allowing for the hereditary use of modules. While some functions are pre-prepared, compressed, and stored, others are calculated during the process to optimize performance. This approach, along with compliance with parallelization and memory management, ensures that the system operates at the highest efficiency. Additionally, the software offers powerful visualization tools that enable users to effectively analyze reconstruction results. Its open-source nature promotes further customization and development by different academic teams, enhancing its utility and impact in the scientific community. Various combinations of different detector and object sizes and positions, source and detector movement, scanning angle, number of iterations and selected projection, back-projection, reconstruction and regulatory methods have been tested to emphasize the impact and importance of each module of the system and to compare effective methods. Reconstruction methods to be examined include SART, MLEM, FDK techniques, and a comparison of these methods has been made. As a result, this modular system simulation software, which is proposed in order to obtain high-quality and reliable images in the low-dose tomographic imaging process, will make significant contributions by providing suitable solutions for both CT and DBT. In addition, given the different parameters used by commercial systems, it also offers a common system simulation that can be used for all these.