Spatiotemporal learning of dynamic positron emission tomography data improves diagnostic accuracy in breast cancer.

Positron emission tomography (PET) tracer distribution is a dynamic process where tissue-specific biochemical properties are reflected in temporal PET dynamics. This cannot be accurately accounted for by conventional static SUV imaging. We posit that PET diagnostic accuracy can be improved exploiting the biochemical and metabolic information embedded in the shape of the tissue time activity curves obtained with a dynamic PET acquisition. We aimed to discriminate tumoral from healthy tissue in a cohort of 88 breast cancer patients who received dynamic $3^{\prime}\mbox{-}{deoxy}\mbox{-}3^{\prime}$ - $^{18}F$-fluorothymidine PET scans with deep-learning filters which learn temporal patterns from 1D time sequences. Compared to the gold standard SUV method ($85%$ accuracy), the best performance was obtained by our CONV1D model which delivered $92%$ accuracy. Our method is superior to conventional SUV analysis and paves the way for more sophisticated applications where deep-learned time signal intensity pattern analysis can be used for tumor segmentation or kinetic assessment without any pharmacokinetic model or invasive measurement of the arterial input function.