Abstract
241
Objectives: [11C]UCB-J is an SV2A PET radioligand that exhibits excellent characteristics including high brain uptake and fast kinetics. To apply this tracer for more applications including pediatric imaging, reducing the injection dose is needed to minimize radiation exposure. Dose reduction would lead to noisier dynamic activity images, resulting in noisy and biased estimation of kinetic parameters. Conventional procedures to reduce noise in images apply post-reconstruction filtering, which has an inherent drawback of resolution loss introducing intensity bias. The goal of this study is to develop an artificial neural network (ANN) based denoising method to suppress noise while maintaining local intensities in multiple dynamic frames reconstructed from reduced-dose PET data. The kinetic parametric images estimated from the ANN processed dynamic frame sequences are evaluated.
Methods: We studied three [11C]UCB-J PET scans of healthy subjects, which were taken in list-mode for 120 min post injection on an HRRT. For each scan, three synthetic 1/3-dose datasets were generated by redistributing the full-dose events sequentially to 3 replicates. Both the full-dose and the 1/3 dose datasets were reconstructed into 33 frames using the MOLAR algorithm. A fully-connected feedforward ANN with an input layer, a hidden layer, and an output layer was trained to form a nonlinear patch-to-patch mapping between the dynamic frames reconstructed from the reduced-dose data and the full-dose data. The training input were 300K 3D image patches of the 30th dynamic frame reconstructed from the 1/3-dose PET data of scan 1, while the training output were the corresponding patches from its full-dose data. The trained ANN was applied to process the 33 1/3-dose dynamic frames from the other two scans. It is noted that we normalized the training input and output patches to have intensity values within the range of [-1, 1] by using the min-max scaling method. Such a normalization ensures the generalization of the ANN trained by one individual frame of a scan to all the dynamic frames of other scans. The full-dose dynamic image frames and the 1/3-dose frames before and after ANN processing were analyzed voxel by voxel with the one-tissue compartment model to estimate parametric maps of K1 and volumes of distribution (VT). We calculated the cross-replicate ensemble normalized standard deviation (EnNSD) in 22 regions of interest (ROIs) of the dynamic frames to compare the 1/3-dose reconstructions with and without ANN processing. The regional mean differences between the pre and post ANN processing image frames were also computed. To assess the effect of ANN processing on the K1 and VT estimation, we evaluated the regional mean values of the ANN processed parametric maps by computing their relative errors (REs) with the full-dose parametric maps as the reference. As a measure of noise in the estimated K1 and VT maps, the EnNSD over the ROIs across the replicates was computed.
Results: For all dynamic image frames of scan 2, the ANN processing achieves EnNSD reduction of averaged 25% among all ROIs. The regional mean differences between pre and post ANN processing are rather small with the largest difference as 3.3%. For the K1 and VT map estimation, the ANN processing results in the averaged noise reduction of 30% and 22% in the ROIs. The REs of regional mean values in the ANN processed K1 and VT are close to 0 with respect to their counterparts from the full dose estimation. Similar results were obtained for scan 3.
Conclusions: An ANN based denoising technique was developed for reduced-dose dynamic brain PET imaging to suppress the noise while preserving the image local mean intensities. The ANN trained with one individual frame of a scan applies well to all the dynamic frames of other scans. Achieving noise reduction and small REs in the estimated K1 and VT maps, the proposed ANN processing demonstrates the potential of reducing injection dose by ~40% in dynamic brain PET imaging.