Performance of deep learning restoration methods for the extraction of particle dynamics in noisy microscopy image sequences

Image-based particle tracking is an essential tool to answer research questions in cell biology and beyond. A major challenge of particle tracking in living systems is that low light exposure is required to avoid phototoxicity and photobleaching. In addition, high-speed imaging used to fully capture particle motion dictates fast image acquisition rates. Short exposure times come at the expense of tracking accuracy. This is generally true for quantitative microscopy approaches and particularly relevant to single molecule tracking where the number of photons emitted from a single chromophore is limited. Image restoration methods based on deep learning dramatically improve the signal-to-noise ratio in low-exposure datasets. However, it is not clear whether images generated by these methods yield accurate quantitative measurements such as diffusion parameters in (single) particle tracking experiments. Here, we evaluate the performance of two popular deep learning denoising software packages for particle tracking, using synthetic datasets and movies of diffusing chromatin as biological examples. With synthetic data, both supervised and unsupervised deep learning restored particle motions with high accuracy in two-dimensional datasets, whereas artifacts were introduced by the denoisers in 3D datasets. Experimentally, we found that, while both supervised and unsupervised approaches improved the number of trackable particles and tracking accuracy, supervised learning generally outperformed the unsupervised approach, as expected. We also highlight that with extremely noisy image sequences, deep learning algorithms produce deceiving artifacts, which underscores the need to carefully evaluate the results. Finally, we address the challenge of selecting hyper-parameters to train convolutional neural networks by implementing a frugal Bayesian optimizer that rapidly explores multidimensional parameter spaces, identifying networks yielding optional particle tracking accuracy. Our study provides quantitative outcome measures of image restoration using deep learning. We anticipate broad application of the approaches presented here to critically evaluate artificial intelligence solutions for quantitative microscopy.