Deep learning-based time-of-flight (ToF) image enhancement of non-ToF PET scans

PURPOSE To improve the quantitative accuracy and diagnostic confidence of PET images reconstructed without time-of-flight (ToF) using deep learning models trained for ToF image enhancement (DL-ToF). METHODS A total of 273 [$^{18}$F]-FDG PET scans were used, including data from 6 centres equipped with GE Discovery MI ToF scanners. PET data were reconstructed using the block-sequential-regularised-expectation-maximisation (BSREM) algorithm with and without ToF. The images were then split into training (n = 208), validation (n = 15), and testing (n = 50) sets. Three DL-ToF models were trained to transform non-ToF BSREM images to their target ToF images with different levels of DL-ToF strength (low, medium, high). The models were objectively evaluated using the testing set based on standardised uptake value (SUV) in 139 identified lesions, and in normal regions of liver and lungs. Three radiologists subjectively rated the models using testing sets based on lesion detectability, diagnostic confidence, and image noise/quality. RESULTS The non-ToF, DL-ToF low, medium, and high methods resulted in - 28 ± 18, - 28 ± 19, - 8 ± 22, and 1.7 ± 24% differences (mean; SD) in the SUV$_{max}$ for the lesions in testing set, compared to ToF-BSREM image. In background lung VOIs, the SUV$_{mean}$ differences were 7 ± 15, 0.6 ± 12, 1 ± 13, and 1 ± 11% respectively. In normal liver, SUV$_{mean}$ differences were 4 ± 5, 0.7 ± 4, 0.8 ± 4, and 0.1 ± 4%. Visual inspection showed that our DL-ToF improved feature sharpness and convergence towards ToF reconstruction. Blinded clinical readings of testing sets for diagnostic confidence (scale 0-5) showed that non-ToF, DL-ToF low, medium, and high, and ToF images scored 3.0, 3.0, 4.1, 3.8, and 3.5 respectively. For this set of images, DL-ToF medium therefore scored highest for diagnostic confidence. CONCLUSION Deep learning-based image enhancement models may provide converged ToF-equivalent image quality without ToF reconstruction. In clinical scoring DL-ToF-enhanced non-ToF images (medium and high) on average scored as high as, or higher than, ToF images. The model is generalisable and hence, could be applied to non-ToF images from BGO-based PET/CT scanners