CNN-Based Cross-Modal Residual Network for Image Synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Current cross-modal image synthesis methods struggle with spatial and structural information in human tissues, resulting in low-quality images with fuzzy edges and poor signal-to-noise ratio.
• Effective synthesis of medical images like PET scans is crucial for diagnosis and research.

Purpose of the Study:

• To develop an advanced cross-modal image synthesis technique that accurately captures spatial and structural information of human tissues.
• To improve the quality and clarity of synthetic medical images, specifically Positron Emission Tomography (PET) images.

Main Methods:

• A novel cross-modal synthesis approach combining generative adversarial networks (GANs) with enhanced residual modules and an attention mechanism.
• Implementation of a multiscale discriminator to improve feature learning and discriminant performance.
• Inclusion of a multilevel structural similarity loss function to enhance contrast preservation.

Main Results:

• The proposed model demonstrated improved performance on the ADNI dataset compared to mainstream algorithms.
• Significant reductions in Mean Absolute Error (MAE) for synthetic PET images were observed.
• Enhancements in Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR) indicate better image quality.

Conclusions:

• The developed method effectively maintains the structural and contrast information of medical images.
• The enhanced GAN model produces synthetic images with superior visual and objective quality, closely resembling genuine prints.
• This technique offers a promising solution for generating high-fidelity synthetic medical images.