PDDG: Medical Image Fusion based on Progressive Dual-Discriminator GAN

When medical imaging modalities are considered individually, they do not contain sufficient information of various aspects of a tissue. Medical image fusion methods have been proposed such that the information deficiency of each medical imaging modality, which is caused by inherent characteristics of imaging modalities, is eliminated and the resultant outputs contain more information than the individual modalities alone. This study proposes a novel method for medical image fusion of MRI and PET images based on progressive dual-discriminator GAN. The progressive part of the method, adds layers to the discriminators and generator in accordance with the resolution of down-scaled source images in each step. It allows for capturing information from source images at different resolutions and various scales, making the training process more stable, robust, and efficient. On the other hand, dual-discriminators participate in an adversarial game with the generator. The generator part of the model generates fused images, one of the discriminators is responsible for distinguishing between generated fused images and real MRI images, while the other discriminator tries to distinguish between generated fused images and PET images. Finally, each single output image of the network will store information about the functionality of PET images, structural characteristics, and texture details of MRI images. In our proposed method, we have achieved high-quality fused images of PET and MRI data. Our method has achieved good results in terms of visual and qualitative assessments. In quantitative evaluations, three metrics have been employed: SSIM, MI, and PSNR, and our method has gained better results in comparison with other algorithms.