Currently, coded aperture γ camera has been extensively utilized in radioactive source location and imaging task. Owing to the advancement of deep learning technology, numerous image reconstruction algorithms have been developed to address the issue of direct convolution algorithm and iterative algorithm for suppressing random noise whilst the convolutional neural network (CNN) algorithm is one of popular candidates, but it has not been fully discussed in existing studies.

This study aims to explore the influence of CNN algorithm, and the size of dataset on the performance of the imaging model for γ camera.

Based on point source imaging process of coded aperture γ camera, Monte Carlo simulation combined with linear random was applied to generate the dataset. The Geant4 software was employed to simulate and encode the aperture imaging process, and the CNN algorithm was utilized to accomplish image reconstruction.

The image reconstruction results show that, for orphan source imaging, the average Contrast-to-Noise Ratio (CNR) of the model using the ⁵⁷Co training set for ⁵⁷Co source location is 75.8, and the average CNR is 24.7 for ¹³⁷Cs source location, while that of the model using the ¹³⁷Cs training set for both locations is 43.8 and 44.3, respectively. With the increase of datasets capacity, the generalization ability of ⁵⁷Co and ⁶⁰Co training model decreases, while the learning effect of ¹³⁷Cs training model increases gradually. When reconstructing seven random ⁶⁰Co sources in the field of view, the CNR of the optimal model is 8.9, and the location of the radioactive source can be clearly identified.

The finding of this study indicates that the capacity and characteristics of the datasets directly impact the learning and generalization capabilities of the CNN model. In high-energy and multi-point imaging scenarios, augmenting the noise in the training set and selecting an appropriate dataset capacity can enhance the effectiveness and accuracy of localizing and reconstructing radioactive sources.