Deep learning based denoising of super-resolution images of protein aggregates

Abstract

Neurodegenerative diseases (NDs) are closely associated with the amyloid aggregation of proteins like Amyloid-, -synuclein, and tau. Understanding the pathogenesis of NDs requires studying the structures and morphological features of these aggregates, which are typically below 100 nm in size. Traditional fluorescence microscopy is limited by the diffraction limit of light (~250 nm). Single-molecule localization microscopy (SMLM) offers a resolution down to ~20 nm, enabling the visualization of these aggregates. However, issues such as non-specific binding (NSB) of fluorophores and background noise degrade the quality of SMLM images. This study presents a U-net-based convolutional neural network (CNN) to denoise SMLM images of protein aggregates. The training dataset includes noise-free super-resolution images of aggregates and their noisy counterparts with non-specific binding signals. Various imaging conditions are simulated to mimic real-world scenarios. The U-net's output is evaluated against ground-truth images for denoising performance. Post-processing techniques further enhance denoised images. The fine-tuned U-net model achieves a validation loss of 0.0042 and low prediction errors of 0.32% and 3.71% in the area and number of aggregates, respectively. This research offers a powerful tool for denoising SMLM images, facilitating accurate characterization of protein aggregate structures and morphological features.