Learning from irregularly sampled data for endomicroscopy super-resolution: a comparative study of sparse and dense...

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Computational Imaging

Background:

• Probe-based confocal laser endomicroscopy (pCLE) provides optical biopsies but generates irregularly sampled data.
• Current pCLE reconstruction uses linear interpolation, which may not be optimal for image quality.
• Convolutional neural networks (CNNs) show potential for improving pCLE image quality.

Purpose of the Study:

• To compare existing pCLE reconstruction methods with deep learning approaches.
• To develop and evaluate novel CNN architectures for reconstructing high-quality pCLE images from irregular data.
• To investigate the effectiveness of embedding Nadaraya-Watson (NW) kernel regression within a CNN framework.

Main Methods:

• Comparison of pCLE reconstruction and super-resolution (SR) methods using irregularly sampled or reconstructed pCLE images.
• Implementation of a novel trainable CNN layer by embedding Nadaraya-Watson (NW) kernel regression.
• Design of deep learning architectures for direct reconstruction from irregularly sampled pCLE data.
• Generation of synthetic sparse pCLE images for methodology evaluation.

Main Results:

• Both dense and sparse CNNs demonstrated superior performance compared to the current clinical reconstruction method.
• Image quality was assessed using peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM).
• The proposed deep learning approaches effectively reconstructed high-quality pCLE images.

Conclusions:

• The study compared sparse and dense CNN approaches for pCLE image reconstruction.
• A trainable generalized NW kernel regression was implemented as a novel sparse approach.
• Synthetic data generation for pCLE SR training was successfully achieved.