Denoising of high-resolution confocal microscopy images of dendritic spines using noise2void algorithm

Abstract

The denoising is an essential pre-processing step for morphological study of dendritic spines as biomedical images contain inevitably noises due to various reason such as imperfection in microscopy, limited low light environment and slow acquisition speed (shutter speed), continuous movement of target, the consequent experimental trade-off between signal-to-noise ratio, and imaging depth. Different algorithms have been proposed in past three decades with varying denoising performances. In recent years, having outperformed all conventional methods, deep learning based models have shown a great promise in image denoising. Most of the deep learning based denoising using supervised learning methods require paired noisy and noiseless/low-noise images. But in reality, it is hard to get ground truth image. In this thesis, Noise2Void, a self-supervised model is introduced to use noisy confocal microscopy images of dendritic spine for model training and denoising and its denoising performance is assessed.