Medical Image Volumetric Arbitrary Scale Super-Resolution Via Learnable Adaptive Upsampling and Weight Dynamic...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Volumetric medical imaging (CT, MRI) often suffers from anisotropic resolution, limiting diagnostic accuracy.
• Existing 2D super-resolution (SR) methods struggle with volumetric data due to anatomical variations and inefficient context aggregation.
• Current arbitrary-scale SR methods present a trade-off between flexibility and parameter efficiency.

Purpose of the Study:

• To develop a robust and parameter-efficient volumetric super-resolution (SR) network for medical imaging.
• To address the challenges of anisotropic resolution and anatomical variations in CT and MRI data.
• To enable high-fidelity SR at arbitrary integer and decimal scales.

Main Methods:

• Proposed a Learnable Adaptive Upsampling-based Volumetric SR network (LAUVSR).
• Introduced a volumetric-anisotropy-driven upsampling core with in-plane weight sharing and through-plane independent generation.
• Incorporated a volume-specific criss-cross attention mechanism and reinforcement-learning-guided weight balancing for improved robustness.

Main Results:

• Demonstrated high-quality SR results across a dense range of integer and decimal upsampling scales on CT and MRI datasets.
• Achieved parameter-efficient SR at arbitrary scales.
• Showcased robustness across datasets, modalities, and unseen upsampling scales.

Conclusions:

• The proposed LAUVSR method effectively enhances volumetric medical image resolution, improving through-plane details.
• The network offers a robust and parameter-efficient solution for arbitrary-scale volumetric SR.
• LAUVSR shows significant potential for improving clinical diagnosis through enhanced medical image quality.