Skip to Main Content
The aim of this study was to improve image quality of statistical reconstruction by using the system matrix (SM) which was trained with artificial neural network (ANN). For training the ANN SM, six input and desired output pairs were generated. The inputs, the digital images, were generated by scanning the mini deluxe cold spot phantom at six different orientations using an optical scanner (resolution: 2400 dpi). Furthermore, the desired outputs were generated by acquiring the projection data with the corresponding angles using the microPET R4. The input and output pairs were used for training the ANN SM. In ANN method, the AD ALINE network with a bias vector was used. The SM, which was created by the geometric information, was treated as the initial guess, and the learning rate was set to le-7 to avoid learning error. The moment term (r = 0.95) was included to improve the stability of the optimization. The inputs were sent to the ANN SM which was updated to minimize the mean square error between the outputs and desired outputs. Moreover, a multi-line source phantom was scanned to obtain the spatial resolutions for comparison. A rat FDG microPET image was acquired to compare the difference between the results reconstructed by the original SM and those by the ANN SM. In the experiment of multi-line source phantom, the resolutions of reconstructed image were measured at center, 10 mm, 20 mm and reconstructed FWHMs by the original SM were 1.61,1.82, and 2.32 mm, respectively. The resolutions reconstructed by the ANN SM were 1.27,1.66, and 1.89 mm. The data size of the original SM and the ANN SM were 90.1,180.0 MB, respectively. The time for reconstruction of the original SM was two times faster than that of the ANN SM. The image quality reconstructed by the ANN SM is better than that reconstructed by the original SM. The results suggested that SM can be updated toward ideal SM using ANN for statistical reconstruction.