Arbitrary Scale Super-Resolution for Medical Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning-based single image super-resolution (SISR) shows promise in medical imaging for enhancing image quality without extra scans.
• Existing SISR methods are often limited to specific magnification scales and lack generalizability.
• There is a need for versatile super-resolution techniques applicable to diverse medical imaging modalities and scales.

Purpose of the Study:

• To develop a medical image arbitrary-scale super-resolution (MIASSR) method capable of generalizing across various magnification scales.
• To leverage meta-learning and generative adversarial networks (GANs) for arbitrary-scale super-resolution in medical images.
• To demonstrate the adaptability of the proposed method to different medical imaging modalities through transfer learning.

Main Methods:

• Coupling meta-learning with generative adversarial networks (GANs) for arbitrary-scale super-resolution.
• Training and evaluating the MIASSR model on single-modal and multi-modal magnetic resonance (MR) brain image datasets (OASIS-brains, BraTS).
• Employing transfer learning to adapt MIASSR for new medical modalities like cardiac MR (ACDC) and chest CT (COVID-CT).

Main Results:

• MIASSR achieves comparable fidelity and superior perceptual quality compared to state-of-the-art SISR algorithms on brain MR images.
• The proposed method demonstrates the smallest model size among evaluated SISR approaches.
• Transfer learning enables MIASSR to effectively perform super-resolution on cardiac MR and chest CT images.

Conclusions:

• MIASSR offers a robust solution for arbitrary-scale super-resolution in medical imaging, outperforming existing methods in perceptual quality and model efficiency.
• The method's adaptability via transfer learning makes it suitable for a wide range of clinical applications.
• MIASSR has the potential to serve as a foundational pre-/post-processing tool in clinical image analysis, improving tasks like reconstruction, enhancement, and segmentation.