Attention-Aware Discrimination for MR-to-CT Image Translation Using Cycle-Consistent Generative Adversarial Networks.

Area of Science:

• Medical Imaging
• Deep Learning
• Artificial Intelligence

Background:

• Magnetic Resonance (MR) and Computed Tomography (CT) are crucial imaging modalities in oncology.
• Accurate image translation between MR and CT is vital for treatment planning and monitoring.
• Current MR-to-CT translation methods face limitations in accuracy and efficiency.

Purpose of the Study:

• To introduce an Attention-aware, Cycle-Consistent Generative Adversarial Network (A-CycleGAN) enhanced with Variational Autoencoding (VAE).
• To establish A-CycleGAN as a superior alternative to existing state-of-the-art MR-to-CT image translation techniques.

Main Methods:

• Incorporated an attention-gating mechanism into the discriminator for parsimonious parameter usage.
• Enhanced the model with VAE to enable deeper discrimination architectures without compromising convergence.
• Trained and validated the A-CycleGAN model using data from 60 head, neck, and brain cancer patients.
• Evaluated performance using Mean Absolute Error (MAE) and Peak Signal-to-Noise Ratio (PSNR) on a holdout test set of 30 patients.

Main Results:

• A-CycleGAN demonstrated superior performance compared to U-Net, Generative Adversarial Network (GAN), and Cycle-Consistent GAN.
• Achieved a mean MAE of 19.61 (95% CI: 18.83, 20.39, P = .0104).
• Obtained a mean Structure Similarity Index Metric (SSIM) of 0.778 (95% CI: 0.758, 0.798, P = .0495) and a mean PSNR of 62.35 (95% CI: 61.80, 62.90, P = .0571).

Conclusions:

• A-CycleGAN represents a significant advancement in MR-to-CT image translation.
• The proposed method offers a superior alternative to current state-of-the-art techniques.
• This technology holds promise for improving oncological imaging analysis and treatment planning.