NExpR: Neural Explicit Representation for fast arbitrary-scale medical image super-resolution

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image rescaling is crucial for interpretation at various resolutions.
• Conventional deep learning super-resolution (SR) is limited to fixed scales.
• Implicit Neural Representation (INR) enables arbitrary-scale SR but is computationally slow.

Purpose of the Study:

• To introduce Neural Explicit Representation (NExpR) for efficient arbitrary-scale medical image SR.
• To overcome the speed limitations of existing INR-based SR methods.
• To achieve high-quality medical image rescaling with significantly reduced processing time.

Main Methods:

• NExpR represents medical images using explicit analytical functions derived from low-resolution inputs.
• A single neural network (NN) inference generates the analytical function parameters.
• Arbitrary-scale SR images are obtained by evaluating these explicit functions.

Main Results:

• NExpR achieves significant speedups, reducing rescaling time by over 100x (from 1 ms to 0.01 ms).
• The method maintains or surpasses image quality compared to existing SR techniques.
• Experiments on diverse datasets (MRI, CT) validate NExpR's effectiveness.

Conclusions:

• NExpR provides a fast and effective solution for arbitrary-scale medical image super-resolution.
• The explicit analytical representation offers a substantial advantage in processing speed.
• NExpR demonstrates strong performance across various medical imaging modalities and datasets.