A Reverse Method of Data Augmentation for High Quality PET Image Synthesis

Positron Emission Tomography (PET) is crucial in oncology for cancer diagnosis and management but involves radiation exposure that has health concerns. An alternative is to use low-dose PET (LPET) instead of standard-dose PET (SPET) where reduced LPET administrates lower radiation dosage fot the scan. However, low-dose PET (LPET) images typically have low signal-to-noise ratio (SNR) and poorer image quality when compared to SPET images, which inadvertently affect diagnostic accuracy. Therefore, using generative artificial intelligence (AI) methods to synthesize (via image generation) SPET images from LPET images has become an active research. Methods based on using convolutional neural networks (CNNs) and Generative Adversarial Networks (GANs) have shown promising performance in SPET synthesis from LPET. However, the performances of these models are highly reliant on the number of paired training samples to learn the complex mappings between LPET and SPET images, e.g., LPET and SPET of the same patient. Without sufficient number of paired training samples to cover all the possible variations, e.g., patient-specific tumour structures, complex anatomical structures, and variety of diseases, existing methods tend tosuffer in generating accurate synthetic SPET. In this study, we propose a new method to boost the number of training samples for LPET to SPET synthesis, where we introduce a new ‘reverse’ data augmentationfor SPET synthesis where we exploit large amounts of SPET images that are acquried as part of a routine imaging. In our approach, we do not need a paired training data, and instead, we convert the SPET images into their LPET image counterparts; these converted images are then used to augment the existing training datasets. Our experiments with the Ultra-low Dose PET Imaging Challenge benchmark datasets show that our method was able to enhance the state-of-the-art methods for producing SPET images, especially for the challenging studies.