Super-resolution biomedical imaging via reference-free statistical implicit neural representation.

Area of Science:

• Biomedical Imaging
• Deep Learning
• Image Processing

Background:

• Supervised deep learning for image super-resolution (SR) faces challenges in biomedical imaging due to the scarcity of paired low-resolution (LR) and high-resolution (HR) images for training.
• Existing methods often require extensive datasets, limiting their applicability in specialized medical imaging scenarios.

Purpose of the Study:

• To develop a reference-free, unsupervised deep learning framework for generating high-quality biomedical SR images.
• To address the limitations of supervised SR methods by utilizing only single or a few observed LR images.

Main Methods:

• A statistical implicit neural representation (INR) framework was proposed, modeling LR image statistics using maximum likelihood estimation.
• An INR network, a coordinate-based multi-layer perceptron, was trained to represent the latent HR image as a continuous spatial function.
• The method ensures functional smoothness and supports arbitrary scaling for SR imaging.

Main Results:

• The framework's effectiveness was validated on diverse biomedical imaging modalities, including CT, MRI, fluorescence microscopy, and ultrasound.
• Successful SR image generation was achieved across various magnification scales (2×, 4×, 8×) using limited LR data.
• The proposed statistical INR approach demonstrated its potential for high-quality SR reconstruction.

Conclusions:

• The developed unsupervised deep learning framework offers a significant advancement for biomedical SR applications lacking HR reference data.
• This reference-free statistical INR method provides a viable solution for numerous SR tasks in medical imaging.
• The approach overcomes data limitations inherent in supervised learning for biomedical SR.