Adaptive Multimodality Translation for Nonattenuation Corrected Total-Body PET

CT and MR images are usually used as an accompaniment for PET to help doctors diagnose diseases. However, it is resource intensive and cannot ensure the consistency of patient status when using PET/CT or PET/MR systems to obtain additional CT or MR images. Thus, it is necessary to use cross-modal transformation to synthesize CT or MR images from PET images. To dynamically generate specified modality images from non-attenuation-corrected PET images according to user requirements, we design an adaptive model translation module to empower the translation model with the ability of adaptively multimodality translation. The adaptive model translation learns the scale and shift parameters from the switch code by two fully connected layers. The scale and shift parameters determine the modality of the model output and are the key to adaptively multimodality translation. The proposed translation can also realize direct attenuation correction by translating nonattenuation correction PET to attenuation-corrected PET, which can reduce patient radiation exposure. Qualitative and quantitative analysis shows that the proposed method performs well with a slight qualification error, good image quality, and high image similarity. This makes the proposed network a promising tool in preclinical research, such as tumor contouring, anatomical localization, and dose calculation of radiotherapy.