Abstract:

Synthesizing magnetic resonance (MR) images for missing contrast is a critical and challenging task. Existing methods usually construct an implicit contrast-shared space with auxiliary loss and modulate the translation process with one-hot contrast code to achieve multi-contrast translation in a single model. We propose an effective method, ProGAN, which is a prompt-driven unified unpaired MR image translation method based on a generative adversarial network (GAN). ProGAN leverages learnable prompts and text embeddings from contrastive language-image pre-training (CLIP) as external knowledge to generate multi-level dynamic prompts, providing powerful and robust guidance on contrasts. Extensive experiments on two public multi-contrast MRI datasets showed that ProGAN achieved state-of-the-art performance both quantitatively and qualitatively.

Key words: multi-contrast magnetic resonance images, unpaired image translation, generative adversarial network, contrastive language-image pre-training (CLIP)