PyTomography: A GPU-accelerated, Opensource Python Library For Image Reconstruction

• Published in: 2023 IEEE Nuclear Science Symposium, Medical Imaging Conference and International Symposium on Room-Temperature Semiconductor Detectors (NSS MIC RTSD) p. 1
• Main Authors: Polson, L., Fedrigo, R., Sabouri, M., Dzikunu, O., Ahamed, S., Rahmim, A., Uribe, C.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.11.2023
• Subjects: Positron emission tomography; Reproducibility of results; Semiconductor detectors; Single photon emission computed tomography; Soft sensors; Software algorithms; Task analysis
• ISSN: 2577-0829
• DOI: 10.1109/NSSMICRTSD49126.2023.10338217

Fast fully-quantitative tomographic image generation opens up new possibilities in research and practice. For different tasks, numerous images may have to be generated, such as for large-scale studies, for dynamic imaging, and/or to select best reconstructions amongst many candidate techniques. As an example, optimal image reconstructions in theranostic imaging as applied to radiopharmaceutical therapies for improved assessments and personalizations remains an important frontier. We introduce PyTomography, a medical imaging tomographic reconstruction framework that generates quantitative Single Photon Emission Computed Tomography (SPECT) and Positron Emission Tomography (PET) images. It is developed in Python and uses the GPUaccelerated functionality of PyTorch to efficiently perform the mathematical operations required for image reconstruction. While the software is straightforward and permits reconstruction from a variety of simulated and real data sources, such as the DICOM format, it also provides the necessary infrastructure and flexibility for the development of novel reconstruction techniques and algorithms. Development in an open-source software framework can enhance reproducibility across vendors and research groups, and may expedite corresponding implementation in a clinical setting. Our research validates PyTomography against vendor-specific reconstructions, explores the incorporation of novel prior functions in SPECT reconstruction, and tests the use of comprehensive resolution modeling in PET imaging.