%0 Journal Article %A Jing Tang %A Zhihui Sun %A Bao Yang %A Wenzhuo Lu %A Chi Liu %T Artificial Neural Network Based Noise Reduction for Chest PET Imaging %D 2019 %J Journal of Nuclear Medicine %P 246-246 %V 60 %N supplement 1 %X 246Objectives: Lower dose in PET imaging is desired but results in higher noise level which makes clinical diagnosis difficult. Therefore, methods to reduce noise in low-dose PET imaging are under study. The recent advances in machine learning technology provide new means for PET noise reduction research. The goal of this study is to apply an artificial neural network (ANN) for image-domain denoising in low-dose chest PET imaging. Methods: Clinical chest PET data of 10 patients with CT detected indeterminate lung nodules were used for this study. Each patient dataset was acquired by a Siemens mCT scanner with 10 mCi FDG injection and the 10 nodules ranged between 5 to 23 mm with 8 of them smaller than 10 mm. For each patient, ten replicates of low-count data with 1/10 of the full counts within 60-80 min post injection were generated by uniform down-sampling of the list mode data. OS-EM reconstructions were then performed for the full-count and low-count data without post filtering. The first patient dataset was used for training the ANN, while the rest 9 patient datasets were used for evaluation. We adopted a 3-layer ANN model with one 128-unit hidden layer in this study. For the training stage, 400K 4×4×4-voxel data patches were selected from one 1/10-count image volume, with the corresponding voxel cubes in the full-count image volume selected as the label patches. After being normalized to be within 0 and 1, these patch pairs were fed into the ANN for training. The stochastic gradient descent algorithm was chosen for loss minimization with a constant learning rate set as 0.01. The training completed after 8 minutes on a 2.7 GHz CPU with 40K iterations. The trained ANN model was evaluated on the ten 1/10-count chest PET image volumes of each of the 9 testing patients. The images were broken down into patches, which were normalized in the same way as was done in the training stage. After applying the trained ANN model, the predicted label patches were assembled back to form the 1/10-count ANN processed images. It required 5 minutes to obtain one 400×400×109 image volume on the same CPU without parallel computing. To quantitatively compare the results without and with ANN processing, noise and nodule signal to noise ratio (SNR) were calculated for each of the tested patients. The noise was calculated as the normalized standard deviation of nodule’s background-voxels around the nodule, while the nodule SNR was calculated as the contrast between the nodule and its background divided by the average standard deviation of the two. The 10 low-count volumes of each patient were treated as 10 noise realizations and the aforementioned standard deviations reflect the variation among them at each voxel location over the regions of interest. Results: For the testing patients, the 1/10-count ANN processed images showed an averaged 40% (38%-43%) noise reduction, compared with their pre-processed counterparts. Meanwhile, the nodule SNR of ANN processed images increased over 40% for all patients. Moreover, in visual examination of both trans-axial and coronal views, the 1/10-count ANN processed images demonstrated significantly lower noise and higher nodule contrast, which is consistent with the quantitative analysis results. Conclusions: The 3-layer ANN model is effective in noise reduction of low-count chest PET imaging. The ANN model trained on one single patient images generalized well and demonstrated promising performance on all other patients tested. With its relatively low computational cost, the ANN provides a practical option for time-sensitive applications. %U