An Arbitrary Scale Super-Resolution Approach for 3D MR Images via Implicit Neural Representation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• High-resolution (HR) medical images are crucial for accurate diagnosis.
• Acquiring isotropic 3D HR MRI is challenging due to long scan times and low signal-to-noise ratio (SNR).
• Existing super-resolution (SR) methods are often limited to fixed up-sampling rates.

Purpose of the Study:

• To develop an Arbitrary Scale Super-Resolution (ArSSR) approach for recovering 3D HR MR images.
• To overcome the limitations of scale-specific SR methods in medical imaging.
• To enable flexible and efficient HR MR image reconstruction.

Main Methods:

• Proposed ArSSR model represents LR and HR images using a shared implicit neural voxel function.
• Employs a convolutional encoder to extract features from LR images and a fully-connected decoder to approximate the implicit function.
• Trained on paired LR and HR MR image datasets to learn the implicit voxel function.

Main Results:

• ArSSR achieves state-of-the-art SR performance for 3D HR MR image reconstruction.
• A single ArSSR model successfully reconstructs HR images at arbitrary up-sampling scales.
• Experimental results on three datasets validate the model's effectiveness and flexibility.

Conclusions:

• ArSSR offers a versatile solution for 3D HR MR image super-resolution.
• The implicit neural representation allows for infinite and arbitrary up-sampling rates.
• This approach enhances the potential for improved diagnostic capabilities through advanced image reconstruction.