Noninvasive Estimation of Macro-parameters for Dynamic PET Images Using Deep learning

In dynamic positron emission tomography (PET), kinetic parameters can quantitatively show the state of the biochemical reaction of the tracer in the body. In general, the values of macro-parameters are more stable than those of micro-parameter, so they are valuable for clinical diagnosis. To estimate the macro-parameters, traditional methods require accurate plasma input functions which are usually obtained by arterial sampling. Unfortunately, collecting arterial blood is not only invasive to the patient, but also difficult to operate. In order to reduce the invasion to patients, we proposed a method to obtain the macro-parameters non-invasively. Deep learning techiques were adopted to learn the nonlinear relationship between the time activity curves (TACs) of tracer and macro-parameters. Our method consisted of two part: training and estimation. In the training part, the deep neural network (DNN) constantly adjusted its own weights to learn the potential functional relation between macro-parameters of interest and dynamic PET images. In the estimation part, desired parametric maps would be acquired directly as long as entering the dynamic PET data into the well-trained DNN. The experiment based on simulation dataset of [18F]-labeled fluorodeoxyglucose was conducted as validation. Our experimental results were compared with those obtained by Patlak plot, which demonstrated the superior performance of the proposed approach.