The estimation of macro-parameters is a critical step in dynamic positron emission tomography (PET) analyses because these parameters can quantitatively characterize the physiologic state of the tracer $in~vivo$ . Numerous algorithms have been proposed to estimate macro-parameters. However, the implementation of these estimation algorithms requires accurate input functions which are difficult to obtain under noninvasive conditions. We developed a novel data-driven framework to estimate the macro-parameter noninvasively without input functions. We took the macro-parameters as a nonlinear function of the activity concentration of the tracer. Deep learning was used to determine this nonlinear function to obtain the macro-parameter directly from dynamic PET data. Our approach was divided into two phases: 1) training and 2) estimation. In the training phase, a deep neural network (DNN) learned the potential relationships between the dynamic PET data and the macro-parameters. In the estimation phase, the macro-parameter could be directly obtained when the dynamic PET data were inputted to the DNN. Experiments based on simulation datasets of 18F-FDG and 11C-FMZ and real datasets were conducted as validation. Our experimental results were compared with those by Patlak or Logan plot, which demonstrated the superior performance of the proposed approach in terms of robustness and accuracy.
Paper
Full text
Noninvasive Estimation of Macro-Parameters by Deep Learning
Semantic Scholar · Medicine · 2020
Abstract
The estimation of macro-parameters is a critical step in dynamic positron emission tomography (PET) analyses because these parameters can quantitatively characterize the physiologic state of the tracer $in~vivo$ . Numerous algorithms have been proposed to estimate macro-parameters. However, the implementation of these estimation algorithms requires accurate input functions which are difficult to obtain under noninvasive conditions. We developed a novel data-driven framework to estimate the macro-parameter noninvasively without input functions. We took the macro-parameters as a nonlinear function of the activity concentration of the tracer. Deep learning was used to determine this nonlinear function to obtain the macro-parameter directly from dynamic PET data. Our approach was divided into two phases: 1) training and 2) estimation. In the training phase, a deep neural network (DNN) learned the potential relationships between the dynamic PET data and the macro-parameters. In the estimation phase, the macro-parameter could be directly obtained when the dynamic PET data were inputted to the DNN. Experiments based on simulation datasets of 18F-FDG and 11C-FMZ and real datasets were conducted as validation. Our experimental results were compared with those by Patlak or Logan plot, which demonstrated the superior performance of the proposed approach in terms of robustness and accuracy.