Ea-GANs: Edge-Aware Generative Adversarial Networks for Cross-Modality MR Image Synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Magnetic resonance (MR) imaging offers diverse tissue contrasts via configurable scanning parameters.
• Cross-modality MR image synthesis leverages complementary information from different imaging types.
• Existing synthesis methods often neglect textural details, impacting synthesized image quality.

Purpose of the Study:

• To propose novel edge-aware generative adversarial networks (Ea-GANs) for improved cross-modality MR image synthesis.
• To address the limitation of existing methods that overlook textural information in image synthesis.
• To enhance the quality and structural integrity of synthesized MR images.

Main Methods:

• Developed edge-aware generative adversarial networks (Ea-GANs) integrating edge information to preserve textural details.
• Proposed two frameworks: generator-induced Ea-GAN (gEa-GAN) and discriminator-induced Ea-GAN (dEa-GAN).
• Utilized 3D-based architecture with hierarchical features for contextual information capture.

Main Results:

• The proposed Ea-GANs, particularly dEa-GAN, significantly outperformed state-of-the-art methods in qualitative and quantitative evaluations.
• dEa-GAN demonstrated superior performance in preserving textural details and structural information.
• The method showed excellent generality, performing well on generic image synthesis tasks beyond MR imaging.

Conclusions:

• Edge-aware generative adversarial networks (Ea-GANs) represent a significant advancement in cross-modality MR image synthesis.
• The integration of edge information is crucial for generating high-fidelity MR images with preserved textural characteristics.
• The dEa-GAN framework offers robust performance and broad applicability in medical and general image synthesis.