Deep Denoising for Scientific Discovery: A Case Study in Electron Microscopy

Feature detection and hypothesis testing for extremely noisy nanoparticle images using topological data analysis

“…. , N there is some subset V k ⊂ I that represents the vacuum, and as such, is purely constituted of Poisson shot noise, as has been assumed in Levin, Lawrence and Crozier (2020)-heuristically verified using plotting heuristics for the vacuum region of Pt nanoparticles in Mohan et al (2022).…”

“…As noted in Lawrence et al (2019), there is a continuing need for algorithms to automate the extraction of relevant features from images to facilitate the assessment of dynamics, and to do so in the presence of an immense amount of noise. Denoising methods based on convolutional neural networks trained on simulated nanoparticle configuration images have been developed specifically to deal with such ultra-noisy nanoparticle videos (Mohan et al, 2022). However, it may not always be feasible to implement these methods in practice, frame-by-frame, on the aforementioned terabytes of data.…”

Machine learning in electron microscopy for advanced nanocharacterization: current developments, available tools and future outlook

“…2 GAN) [22][23][24] or case-independent denoising models that successfully outperformed classical restoration filters for both TEM and STEM. 23,[25][26][27][28][29][30][31] Interestingly J. Vincent et al studied the latent features learned by the DL model to unveil the nature of the trained denoising dependencies to shine light on what is typically left as a black box. They showed that the FCNN learns to adapt its filtering strategies depending on the structural properties of every particular region in the image.…”

“…[92][93][94][95] Currently, the main trend of DL in the field has been the finding of atomic positions and their correlation with mathematical graphs, as stated before, for defect identification [59][60][61]66,67,70 and quantification, 96 or image denoising. 21,[26][27][28] Nevertheless, the extraction of further properties with physical meaning can be envisioned from this idea, such as atom counting or quantitative TEM in a broader sense. 3,4,66 Interestingly, the atomic positions identification started with FCNN on 2D systems and model 3D systems.…”

A Framework for the Optimal Selection for High-Throughput Data Collection Workflows by Autonomous Experimentation Systems

“…taking images faster with a shorter dwell time) decreases image quality and might therefore lead to greater deviations from a specified M . Recent advances have shown that deep-learning algorithms can be used to boost the SNR of low-quality microscopy images and yield images with effectively higher SNR [ [26], [27], [28], [29], [30], [31], [32]]. These methods have exhibited robust performance on images outside of the testing dataset when compared to more standard denoising techniques.…”