Deep learning based super-resolution for large field of view imaging of the porosity network in dentin

Abstract

Imaging dentinal porosity is a challenging topic in dental research. This porosity, consisting of microscopic tubules interconnected with sub-microscopic branches, houses odontoblast cellular processes that bathe in physiological fluids, which are believed to play a key role in mechano-sensing. The recent observation that porosity forms a dense network led to the realization that stimuli propagation could be more complex than currently thought. A 3D representation of this network is therefore key to understanding tooth function. This imposes strong constraints on required imaging resolution (∼ 100 nm) and field-of-view (FOV) to visualize porosity of an entire tooth section, currently beyond our reach. To achieve large-scale high-resolution (HR) visualization, we propose using deep learning (DL) super-resolution (SR) models trained on confocal fluorescence microscopy images, to restore HR details from low-resolution (LR) images acquired much faster by under-sampling. Four DL models (RCAN, FSRCNN, pix2pix, CycleGAN) were trained on experimentally acquired HR/LR pairs, with pixel size increase of x2, x4, and x8. Quantitative analysis was performed using standard image quality assessment metrics, which produced contradictory results compared to visual assessment. This drew the need for the development of a biology-driven evaluation based on 2D connected component and 3D graph network analysis, allowing for better interpretation of performance specific to porosity features. CycleGAN and pix2pix performed best, up to x8, which decreases scan-time by a factor of 20.3. CycleGAN was selected and improved with new training on an enriched dataset for application on large FOV LR acquisitions. Overall results showed great promise for the use of SR models to restore HR information from LR acquisitions. With this approach, including scan-time and model application, a large HR FOV could be generated 8.1x faster than standard HR acquisitions, saving over 300 hours. This introduces potential for large-scale HR visualization of dentin porosity, working towards full tooth visualization.