Performance improvement of gamma-imaging system by selective fusion of deep-learning-based collimator-less and Compton images with FlowNet model: A Monte-Carlo simulation

Collimator-less imaging systems based on deep learning (DL) have been proposed as rapid and computationally efficient alternatives that can overcome field of view (FOV) constraints and eliminate the need for post-reconstruction algorithms of conventional gamma-imaging systems. However, DL models are prone to prediction errors, including hallucinated structures, owing to their dependence on training distributions. These limitations can be mitigated by incorporating information from conventional imaging systems whose reconstruction processes are grounded in physical interaction models that preserve spatial and spectral information. In this study, we propose a FlowNet model, originally developed for optical-flow estimation, to fuse DL-based collimator-less images with Compton images. The FlowNet architecture is expected to learn spatial mapping that selectively corrects erroneous or hallucinated predictions from the DL model by referencing the physically grounded Compton image, while simultaneously suppressing the background artifacts inherent to Compton reconstruction. Quantitative evaluation using the structural similarity index measure (SSIM) demonstrated that the FlowNet-based fusion approach enhanced the overall performance of the DL-based collimator-less imaging system. The results verified the feasibility of applying the FlowNet model for multimodal image fusion to compensate for the respective limitations of each modality and improve the robustness and accuracy of gamma-ray source reconstruction.