Positron Emission Tomography (PET) is a key diagnostic tool in oncology, cardiology and neurology. However, the inherent noise and sparsity of the acquisition process pose significant challenges for image reconstruction. State-of-the-art reconstruction methods often produce low-resolution images and struggle to preserve small details. This thesis studies and develops new reconstruction methods based on the Learned Primal-Dual (LPD) reconstruction and enhanced by the Cross-Attention mechanism. A new synthetic generator capable of producing a wide variety of shapes was implemented, along with a new loss function, leading to improvements in both metrics and generalisation power. Using these new elements, four different LPD architectures incorporating Cross-Attention were tested, achieving comparable performance to previous implementations. Although the Cross-Attention mechanism did not significantly improve the LPD reconstruction algorithm, the results suggest its potential in effectively integrating different information and are promising for future applications.