MedSyn: Medical Image Synthesis of Brain Magnetic Resonance Images Based on Hybrid Information Aggregation

Magnetic resonance imaging (MRI) is widely used in clinical scenarios. However, acquisition of certain image modalities may be in limitation due to multiple considerations such as time duration and labor costs. Hence, medical image synthesis can resolve those limitations by estimating the desired target image modality without inducing an actual scan. In this paper, we proposed a novel method which aggregated the hybrid information of MRI to improve the performance of medical image synthesis and addressed it as MedSyn. Specifically, the MedSyn adopted the GAN framework, and made some great improvement. In generator, we proposed a simple yet efficient Tri-Focus Visual Center (TVC) to activate contextual features by establishing inter-dimensional dependencies with the assistance of rotation operations and gaussian functions. Furthermore, we proposed a Hybrid Global-Local Feature Extractor (HGL) module to capture long-range dependencies and exploit potential relationships between spatial and channel dimensions to extract global-local information. Benefitting from the modules of TVC and HGL jointly, the generator could capture obvious image features and underlying semantic information. In discriminator, we proposed a Dynamic Multi-Scale (DMS) Discriminator which could enhance discriminative ability by combining features of different scales dynamically. To evaluate the performance of the proposed method, we conducted experiments on three brain datasets (the Meningioma dataset, the IXI dataset and BRATS dataset), and compared the results with some state-of-the-art methods. The experimental results showed that our MedSyn outperformed other state-of-the-art methods in terms of both quantitative evaluation and qualitative visual assessments.