Brain PET dose reduction using a shallow artificial neural network

Objectives: Dose reduction in PET imaging is desired but leads to noisier reconstruction using the maximum likelihood (ML) algorithm. To reduce noise, maximum a posteriori (MAP) algorithms have been designed to make use of the statistical properties of the raw count data and prior information in the image domain. Although MAP reconstruction achieved positive results in improving reduced dose PET images, they are computationally expensive. The goal of this study is to develop a learning-based denoising scheme in the image domain to directly process 3D image patches from the reconstructions with reduced counts. To suppress image noise at different count levels while maintaining the local mean intensities, we designed a shallow artificial neural network (ANN) serving as a nonlinear regression function which maps image patches from the reduced-count to the full-count reconstruction.

Methods: A shallow ANN with an input layer, a hidden layer and an output layer was designed to model the nonlinear relationships between the input and the output, which were the image patches from the low-count PET images at different count levels and those from the full-count image. We performed ML reconstructions from the synaptic density tracer ¹¹C-UCB-J from 5 healthy control subjects scanned on the HRRT. For each subject, a full-count brain PET reconstruction, 19 low-count reconstructions (3 with 1/3 of the counts, 6 with 1/6 of the counts, and 10 using 1/10 of the counts) were executed for the period of 40-60 min post-injection. A total of 480K 3D training image patches of size 4x4x4 voxels were selected from the 19 low-count images. Each selected image patch and the label patch from the full-count image were normalized to have intensity values within the range of [0, 1]. Such a normalization ensures the generalization capability of the ANN model on the reconstructed images of other patients. Using the stochastic gradient descent method, the weights of the ANN were updated iteratively to minimize the Euclidian distance between the training input patches and their labels. After training the ANN with 48K iterations, requiring 3 min on a 3.2 GHz CPU, the network was tested to process the low-count reconstructions of the other 4 patients. We normalized each test image and fed the extracted patches to the ANN. It required only 0.7 microsec to process one patch on the same CPU. To evaluate the performance of the proposed ANN processing, we calculated the tradeoff between mean and normalized standard deviation (NSD) in regions of interest (ROIs) of each tested brain image. Moreover, the ANN-processed images were compared with their ML-reconstructed counterparts by calculating the difference and the relative error in the ROIs with respect to the full-count reconstruction.

Results: For every testing image from different count levels and different subjects, the ANN result demonstrates lower NSD and similar regional mean value compared with the corresponding ML reconstruction. More appreciable reduction of image noise is achieved by the ANN processed images for the 1/10 count reconstruction compared with the 1/3 count case. The mean values of the relative error calculated in the image ROIs before and after ANN processing are close to 0. The regional standard deviations (SDs) of the relative error obtained by the ANN processing get an averaged 30% decrease from those of the ML reconstructions at both 1/3 and 1/10 count levels.

Conclusions: An ANN processing technique was developed for brain PET imaging with reduced counts to suppress the image noise while keeping the local mean intensities of the ML reconstruction. The shallow ANN trained with image patches from ML reconstructions at 3 count levels of one subject applies well to image patches of 4 other subjects at different count levels. With robust performance, the proposed ANN processing demonstrated potential for the dose reduction of PET imaging without the necessity of accessing the raw data or vast computational cost.