This work package (WP2) focuses on developing accurate Monte Carlo (MC) radiation transport models to support hardware design and clinical treatment adaptation. To overcome the prohibitive computational costs of traditional simulations, Generative Adversarial Networks (GANs) will be utilized to model secondary radiation distributions and detector responses orders of magnitude faster. By uniquely incorporating timing information as a phase space dimension, these AI-enhanced models will enable real-time applications and rapid adaptation to changes in patient geometry, facilitating the practical clinical use of the NOVO imaging system.

This work package also focuses on solving the complex inverse problem of image reconstruction using a weighted generalized Radon transform over cones. To handle low count rates and missing angles, the project employs advanced regularization techniques and fast optimization solvers pioneered by the University of Manchester. The approach is split into two sub-domains: the detector sub-domain, which utilizes Bayesian Physics Informed Neural Networks (BPINN) to improve event parameter estimates like depth of interaction and time-of-flight; and the image sub-domain, which uses Deep Image Prior (DIP) networks for real-time image refinement and enhanced quality. These methods, combined with library-based spectroscopy, enable 3D spectroscopic imaging and detailed tissue compositional analysis.