Fast Automated PET Image Quality Assessment by Deep Learning

The use of simultaneous positron emission tomography - magnetic resonance imaging (PET-MR) earlier in the diagnostic pathway for memory clinic patients is inhibited by associated radiation dose and tracer costs. Low-dose PET imaging and reconstruction may overcome these limitations, but the higher noise levels compromise image quality. We have developed a convolutional neural network (CNN) capable of automating low-dose PET image quality assessments, typically executed by experienced clinicians. Automating image quality assessments may aid large-scale reconstruction hyperparameter investigations, working faster towards clinically feasible reconstructions at ultra-reduced injected doses. Preliminary work used 1800 reconstructed phantoms to predict, initially, the coefficient of variation (CV) as a noise level measure. For non-normalised and normalised test images, the mean errors in CV prediction were 3.4 ± 2.6% and 16.1 ± 11.9%, respectively. In a real patient study, data from one patient was resampled at different dose levels and reconstructed using maximum-likelihood expectation maximisation. A CNN was trained on 3200 randomly extracted patches, with simultaneous predictions of the dose and three clinician-scored metrics: global quality rating, pattern recognition, and diagnostic confidence. The median prediction for 1000 test-time patches for each dose level was taken, obtaining mean-absolute-errors of 10.20 ± 22.30 MBq, 0.27 ± 0.52, 0.11 ± 0.25 and 0.04 ± 0.04, for each metric, respectively. This study shows that it is possible to automatically predict multiple metrics, both clinician-scored and conventional quantitative image quality measures, simultaneously. Future work will include obtaining more clinically scored reconstructions, to improve training and performance of the model.