Improve DOI resolution for a human brain PET detector using convolutional neural networks

A novel detector design was proposed for high sensitivity and high spatial resolution human brain PET imaging in our previous work, and the detector had the ability to decode the depth-ofinteraction(DOI). To further improve the DOI resolution of the detector, we used NN-based algorithm to predict the DOI for the gamma ray events interacted at different depths of the detector. The detector consist of two layers of LYSO crystal array, top layer was a $16 \times 16$ array of LYSO crystals and bottom layer was an $8 \times 8$ LYSO array. Placed between the top and bottom layers of the LYSO array was a 2mm thick glass light guide for light sharing. The dataset of 15 different depths from DOI $=8.5-22.5 \mathrm{~mm}$ with a step size of 1 mm were acquired and input into the neural network for training, and a separate dataset at three depths of 10, 14, 19mm was acquired and input to the trained neural network for evaluating the DOI prediction accuracy. The DOI resolution was calculated for all crystals within a detector block at three DOI of 10, 14, 19 mm. A comparison was made between the DOI resolution obtained using the w-value method and the NN-based approach, and the average DOI resolution were 4.95 mm and 2.60 mm using the w-value and CNN, respectively. The proposed network provides a 1.9 times improvement in DOI resolution compared with the w-value method. Compared with other NN-based DOI estimation methods, our proposed network was trained with events from all crystals in a detector at multiple depths, instead of individual crystal events. Furthermore, the training process needs fewer epochs and has shown greater improvement on DOI resolution. Preliminary experimental results demonstrated the feasibility of the CNN method in estimating the DOI.