A Novel Method to Improve Image Quality for 2-D Small Animal PET Reconstruction by Correcting a Monte Carlo-Simulated System Matrix Using an Artificial Neural Network

The aim of this study was to improve image quality of statistical reconstruction by using the 2-D system matrix (SM) trained with an artificial neural network (ANN). For training the ANN SM (SMANN), the inputs of ANN, the digital images, were generated by scanning the mini-deluxe cold spot phantom at ten different orientations using an optical scanner (resolution: 0.01 mm/pixel). The desired outputs were generated by acquiring the projection data with the corresponding angles using the micro positron emission tomography (microPET) R4. In the ANN method, the ADALINE network with a bias vector and a momentum term were used. Moreover, a multiline-source phantom and a four-segment phantom were scanned to obtain the spatial resolutions and the quantitative accuracy for comparison, respectively. A rat FDG microPET image was acquired to compare the difference between the results reconstructed by the microPET's built-in 2-D-ordered subsets expectation maximization algorithm (OSEM), OSEMmicroPET, OSEM by Monte Carlo simulated SM, OSEMSMd, and OSEM by the SMANN, OSEMANN. In the multiline-source experiment, the resolutions of OSEMmicroPET measured at center, 10, and 20 mm from the center were 1.61, 1.78, and 2.30 mm, respectively. The resolutions of OSEMSMd and OSEMANN were 1.24, 1.68, and 1.87, and 1.28, 1.62, and 1.72 mm, respectively. In the results of the four-segment phantom, the sum of absolute error of the truth versus the values reconstructed by OSEMmicroPET, OSEMSMd, and OSEMANN were 1085.53, 913.48, and 435.02, respectively. By interpreting the results of the evaluation, the image quality reconstructed by the SMANN is better than that reconstructed by the original SM. The results indicated that SM can be updated toward ideal SM using ANN for statistical reconstruction.