Fully Convolutional Transformer-Based GAN for Cross-Modality CT to PET Image Synthesis

Positron emission tomography (PET) imaging is widely used for staging and monitoring the treatment of lung cancer, but the expensive cost of PET imaging equipment and the numerous contraindications for the examination present significant challenges to individuals and institutions seeking PET scans. Cross-modality image synthesis could alleviate this problem, but existing method still have deficiencies. Such as, pix2pix mode has stringent data requirements, while cycleGAN mode, although it can address this issue, does not produce a unique optimal solution. Additionally, models with convolutional neural network backbone still exhibit limitations when dealing with medical images containing contextual relationships between healthy and pathological tissues. In this paper, we propose a generative adversarial network (GAN) method based on a fully convolutional transformer and residual blocks called C2P-GAN for cross-modality synthesis of PET images from CT images. It composed of a generator and a discriminator that compete with each other, as well as a registration network that can eliminate noise interference. The generator integrates convolutional networks that excel in capturing local image features with the transformer that is sensitive to global contextual information. In the current dataset of 23 pairs of lung cancer patients collected, quantitative and qualitative experimental results demonstrate the superiority of the proposed method relative to competing methods and have great potential for clinical applications.