Cross-Modal PET Synthesis Method Based on Improved Edge-Aware Generative Adversarial Network

Abstract

Current cross-modal synthesis techniques for medical imaging have limits in their ability to accurately capture the structural information of human tissue, leading to problems such edge information loss and poor signal-to-noise ratio in the generated images. In order to synthesize PET pictures from Magnetic Resonance (MR) images, a novel approach for cross-modal synthesis of medical images is thus suggested. The foundation of this approach is an enhanced Edge-aware Generative Adversarial Network (Ea-GAN), which integrates an edge detector into the GAN framework to better capture local texture and edge information in the pictures. The Convolutional Block Attention Module (CBAM) is added in the generator portion of the GAN to prioritize important characteristics in the pictures. In order to improve the Ea-GAN discriminator, its receptive field is shrunk to concentrate more on the tiny features of brain tissue in the pictures, boosting the generator’s performance. The edge loss between actual PET pictures and synthetic PET images is also included into the algorithm’s loss function, further enhancing the generator’s performance. The suggested PET image synthesis algorithm, which is based on the enhanced Ea-GAN, outperforms different current approaches in terms of both quantitative and qualitative assessments, according to experimental findings. The architecture of the brain tissue are effectively preserved in the synthetic PET pictures, which also aesthetically nearly resemble genuine images.