A 4π-field of view deep-learning-based collimator-less imaging system was designed with the Monte Carlo method and performance of the system was studied to verify the feasibility of system. A 4 × 4 × 4 voxelated single-volume GAGG(Ce) system and 57Co, 133Ba, 22Na, and 137Cs point sources at 2000 positions were modeled using Monte-Carlo N-particle transport code version 6 (MCNP6). Two types of the localized energy deposition acquired with a voxelated detector system with and without energy bins, were calculated. The F6 tally was used to provide the entire energy deposited in each voxel and the F8 tally to provide energy spectrum data for each voxel. This system utilized these energy deposition patterns depending on the source type and position to reconstruct the source distribution image. A fully convolutional network which is advantageous for the prediction of image outputs was used to estimate source distribution. The models utilizing energy deposition patterns generated on total energy deposition and energy spectrum data were trained with labels from 30° to 10 degree of full-width half-maximum (FWHM). As a result of training with single and multiple source data, types of isotopes and source locations were discriminated up to 5 sources when using energy spectral data, and the average image similarity between ground truth images and predicted ones were 0.9936 for total energy deposition model and 0.9966 for divided energy bin model. These results showed the feasibility of a collimator-less imaging system based on deep learning method that requires no filtration of any type of interaction.
Read full abstract