Cycle consistent twin energy-based models for image-to-image translation.

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Domain shift, a change in data distribution between training and testing sets, significantly impacts medical image analysis performance.
• Existing generative adversarial network (GAN)-based methods for medical image translation struggle with training stability and output diversity.
• Energy-Based Models (EBMs) present a promising alternative for generative tasks, offering potential advantages over GANs.

Purpose of the Study:

• To introduce a novel method, Cycle Consistent Twin EBMs (CCT-EBMs), for unpaired medical image-to-image translation.
• To address the limitations of current generative models in handling domain shift in medical imaging.
• To improve the accuracy and efficiency of medical image translation tasks.

Main Methods:

• Proposed CCT-EBMs utilize a pair of EBMs within an Auto-Encoder's latent space for bidirectional image translation.
• A novel consistency loss function enforces translation symmetry and domain coupling.
• Theoretical analysis demonstrates reduced Langevin mixing steps for improved efficiency.

Main Results:

• CCT-EBMs achieve superior performance in unpaired medical image-to-image translation compared to state-of-the-art methods.
• Quantitative and qualitative experiments on three diverse medical image datasets validate the method's efficacy.
• The proposed approach demonstrates improved translation quality and domain consistency.

Conclusions:

• CCT-EBMs provide an effective solution for unpaired medical image translation, mitigating domain shift challenges.
• The method offers a more stable and diverse alternative to existing GAN-based approaches.
• This work advances the application of EBMs in medical image analysis, paving the way for improved diagnostic tools.