Global-Token U-Net with Hybrid Loss for Trustworthy Medical Image Super-Resolution

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Super-resolution (SR) technology enhances low-resolution medical images to ultra-high resolutions.
• AI-based SR shows promise in reconstruction quality but often lacks trustworthiness.
• Ensuring reliability and avoiding diagnostic misinformation is critical for medical image SR.

Purpose of the Study:

• To develop a trustworthy super-resolution method for medical images.
• To address the ill-posed nature of SR and ensure diagnostic reliability.
• To enhance high-frequency texture generation while minimizing deviation from ground truth.

Main Methods:

• Designed a novel hybrid loss function combining a hinge-based adversarial term and a PSNR-based regularization term.
• Applied the hybrid loss within a global-token U-Net backbone network.
• Integrated a lightweight VGG network as a discriminator for adversarial training.

Main Results:

• The proposed hybrid loss enhances the trustworthiness of medical image super-resolution.
• The method maintains high reconstruction quality alongside improved reliability.
• Empirical verification confirms the effectiveness of the integrated approach.

Conclusions:

• The novel hybrid loss function improves the trustworthiness of AI-based medical image super-resolution.
• This approach balances the need for refined textures with diagnostic accuracy.
• The method offers a reliable solution for enhancing medical image quality.