Generative Adversarial Networks for Neuroimage Translation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Medical image synthesis is crucial for augmenting limited datasets, enhancing model robustness and generalization.
• Generating diverse neuroimaging data is essential for training reliable diagnostic models.
• Image-to-image translation techniques offer a promising approach for medical data augmentation.

Purpose of the Study:

• To develop and evaluate a CycleGAN model for translating neuroimages between different magnetic field strengths (e.g., 3T to 1.5T).
• To compare the performance of CycleGAN against a deep convolutional GAN for neuroimage translation.
• To assess the accuracy and effectiveness of CycleGAN in generating synthetic and reconstructed neuroimages.

Main Methods:

• Implementation of a cycle-consistent generative adversarial network (CycleGAN) for neuroimage domain translation.
• Comparison of CycleGAN with a deep convolutional GAN architecture.
• Quantitative evaluation using peak signal-to-noise ratio (PSNR) and mean absolute error (MAE).

Main Results:

• CycleGAN successfully generated synthetic and reconstructed neuroimages with reasonable accuracy.
• The mapping from 3T to 1.5T domain achieved an average PSNR of 25.69 ± 2.49 dB.
• The model achieved an average MAE of 2106.27 ± 1218.37 for the translation task.
• CycleGAN outperformed the deep convolutional GAN in generating high-fidelity images.

Conclusions:

• CycleGAN is an effective method for translating neuroimages between different field strengths, aiding in dataset augmentation.
• The developed model contributes to data-oriented robustness by exposing models to diverse visual data.
• Publicly available code facilitates further research and application of this technique in medical imaging.