Experimentally paired high- and low-resolution confocal fluorescence microscopy dataset for deep-learning super-resolution imaging of tooth dentin porosity

TL;DR

This paper introduces a dataset of high- and low-resolution confocal images of tooth dentin porosity for training deep-learning super-resolution models.

Contribution

The dataset provides experimentally paired images and classified patches for super-resolution training in dentin imaging.

Findings

High- and low-resolution images were acquired at multiple pixel sizes for the same field of view.

Images were registered, patched, and classified into three porosity types for training purposes.

The dataset enables both paired and unpaired super-resolution model training with real-world examples.

Abstract

The proposed dataset provides experimentally acquired high- and low-resolution confocal laser scanning microscopy images of dentin porosity, split into classified image patches that can be used for paired or unpaired super-resolution training. Porous dentinal tubule and branch channels represent a vast network and can be used as a model for the mechanosensory system of teeth. The large quantity and small size of tubules and branches present an imaging challenge, requiring high resolution images at a large field of view in order to image statistically relevant regions of the tooth and fully model the porosity network. Imaging more quickly increases the potential of the field of view that can be acquired, however in return decreases resolution. Deep-learning based super-resolution could be used on low-resolution images to restore high resolution information from the image. A fluorescently stained tooth slice was first imaged with 100×100 nm2 pixel size to acquire data at the Nyquist resolution based on the theoretical point spread function of the microscope (216 nm), and low-resolutions images of the same field of view were acquired at 200×200 nm2, 400×400 nm2, and 800×800 nm2 pixel sizes without changing the microscope objective to ensure image region matching. A total of six regions were imaged at 5.1 µm in depth for a total of 18 slices per region. Acquired images were registered to ensure data were paired, and images were further patched and classified as one of 3 classes based on the type and scale of porosity present in each patch. This provides a unique database of classified image patches that can be used for paired or unpaired super-resolution training, providing real examples in all resolutions.

Figures