Neural networks for Gamma cameras

In this research the effectiveness of analytical neural networks compared to the maximum likelihood method on the prediction of spatial and DOI positioning of a Gamma detector with a NaI(Tl) scintillator of size 590mm x 470mm x 40mm (x,y,z), with a glass lightguide of size 620mm x 500mm x 4mm and a PMT area of 620mm x 500mm x 40mm with 2-inch round PMTs with a Bialkali photocathode is presented. This is done by training neural networks with different cost function, different amounts of hidden layers and different amounts of neurons per hidden layer, trained on different amounts of training data. The resolution of the predictions of the testing data are compared with those of the maximum likelihood method. It was concluded that the neural network with best spatial resolution, had the Huber loss function as cost function, 4 hidden layers and 512 neurons per hidden layer and was trained on 29,970 datapoints. The FWHM and the FWTM were 3.83 ± 0.54 mm and 12.49 ± 1.19 mm respectively, while the FWHM and the FWTM of the maximum likelihood method were 3.31 mm and 12.13 mm respectively. The resolution of the neural network was lower than that of the maximum likelihood method. The same was done for the DOI resolution, here a neural network with mean squared error as cost function, 4 hidden layers and 64 neurons per hidden layer trained on 9,990 datapoints, gave the best the resolution with FWHM and FWTM equal to 6.00 ±0.50 mm and 11.94 ± 0.94 mm respectively. The FWHM and FWTM of the maximum likelihood method were 6.16 mm and 11.22 mm respectively. This made the DOI resolution of the neural network higher then that of the maximum likelihood method. Finally different ideas were presented to increase the resolution of the neural network. These were: training the neural network on independent data, split the neural network in a spatial part and a DOI part, create a more complex architecture and making use of a convolutional neural network.