Classification of Neurotransmitter Response in Dynamic PET Data Using Machine Learning Approaches

Dynamic positron emission tomography (PET) imaging can be used to quantify changes in synaptic concentrations of endogenous neurotransmitters during cognitive tasks or pharmacological interventions. Existing pharmacokinetic models, such as the linear parametric neurotransmitter PET (lp-ntPET) method, can be used to model the measured dynamic data and characterize small transient changes in neurotransmitter levels. Application of these models to the voxel level is challenging, however, due to the high levels of noise in dynamic data, leading to a high number of false-positive responses (i.e., low specificity). In this article, we investigated the suitability of machine learning algorithms (MLAs), including support vector machine (SVM) classifiers, shallow feedforward neural networks, convolutional neural networks (CNNs) and long short-term memory (LSTM) networks, to detect and classify transient changes in voxel-wise time-activity curves. We also investigated whether the reconstruction framework (post versus direct reconstruction) had any impact on the performance of the MLAs. We used computer simulations to generate dynamic PET data, representing a [¹¹C]raclopride study, with known activation responses, across a wide range of noise levels. Different simulated data sets were used to train and test the MLAs across a range of noise levels and activation response magnitudes. Results showed the MLAs offered a large improvement in specificity without a corresponding decrease in sensitivity across all noise levels tested compared to direct application of the lp-ntPET model. They also offered a modest benefit over the currently accepted method (statistical F-test combined with cluster size analysis), for both 2-D+time data when incorporated within direct or post-reconstruction frameworks, and 4-D GATE data.