Medical image translation is a process of converting one imaging modality to another, aiming to reduce the necessity for multiple image acquisitions and enhance the efficiency of treatment planning. In this study, we introduce a Generative Adversarial Network (GAN)-based framework for 3D medical image translation, using a 3D multi-resolution Dense-Attention UNet (3D-mDAUNet) as the generator and a 3D multi-resolution UNet as the discriminator. Our model is optimized with a unique combination of loss functions, including voxel loss, voxel-wise relativistic GAN loss, and 2.5D perception loss. This approach yields promising results in volumetric image quality assessment (IQA) across a variety of imaging modalities, body regions, and age groups. Furthermore, we propose a synthetic-to-real applicability assessment as an additional evaluation to assess the effectiveness of synthetic data in downstream applications. Overall, our results not only indicate that our framework produces high-quality synthetic medical images but also demonstrates the potential utility of synthetic data, with extensive testing across various experimental settings.
The figures below present source to target image translation. Left image is a source used as an input to our model. Right image is a translated output image generated by our model.
The figures below present the zoomed comparison among different methods.
The figures below present the holistic comparison among different methods.