Deep Learning methodologies for direct image reconstruction and integrated attenuation correction in brain PET/MRI

This study proposes the ATTDeepPET model, a novel deep learning architecture crafted specifically for advancing positron emission tomography (PET) image reconstruction in PET/MR scanners. By incorporating magnetic resonance (MR) images into its learning process, ATTDeepPET addresses the persistent challenges associated with attenuation effects in PET/MR scanners, eliminating the need for simulated transmission scans. ATTDeepPET's performance is assessed alongside the deep learning model DeepPET, as well as established methods such as FBP, ML-EM, and ML-EMR for comparison. The findings reveal noteworthy achievements since ATTDeepPET accomplishes competitive image quality compared to FBP, MLEM, and ML-EMR when applied to brain phantoms while also demonstrating a reduction in reconstruction times. Nevertheless, when dealing with real PET images, ATTDeepPET does exhibit some performance variability, underscoring the increased complexity of real-set scenarios and the importance of employing diverse datasets to enhance its robustness. Moreover, ATTDeepPET, despite inherent limitations, including heightened memory requirements and sensitivity to dataset variations, presents a promising path forward for PET image reconstruction. Its hallmark traits include exceptional execution speed, liberation from the prerequisite of prior physics knowledge, and the prospect of obviating the need for an additional CT scan for attenuation correction. These attributes hold transformative potential in terms of enhancing diagnostic precision and curtailing patient radiation exposure.