Denoising diffusion models for high-resolution microscopy image restoration

TL;DR

A denoising diffusion probabilistic model is trained to predict high-resolution images by conditioning the model on low-resolution information, and it is shown that this model achieves a performance that is better or similar to the previously best-performing methods, across four highly diverse datasets.

Abstract

Advances in microscopy imaging enable researchers to visualize structures at the nanoscale level thereby unraveling intricate details of biological organization. However, challenges such as image noise, photobleaching of fluorophores, and low tolerability of biological samples to high light doses remain, restricting temporal resolutions and experiment durations. Reduced laser doses enable longer measurements at the cost of lower resolution and increased noise, which hinders accurate downstream analyses. Here we train a denoising diffusion probabilistic model (DDPM) to predict high-resolution images by conditioning the model on low-resolution information. Additionally, the probabilistic aspect of the DDPM allows for repeated generation of images that tend to further increase the signal-to-noise ratio. We show that our model achieves a performance that is better or similar to the previously best-performing methods, across four highly diverse datasets. Importantly, while any of the previous methods show competitive performance for some, but not all datasets, our method consistently achieves high performance across all four data sets, suggesting high generalizability.

Figures (8)

• Figure 1: Conditioned DDPMs outperform several previous methods in denoising STED images. A) Performance comparison based on several evaluation metrics between our proposed method (DDPM and DDPM-avg) and several previously proposed benchmark models (see Methods for description of models and metrics). We indicate the median of the best-performing method for each metric as a dashed line in the respective color. Mood's median test was used to compute statistical significance; ***: $p < .001$, **: $p < .01$, *: $p < .05$, otherwise not significant (top (resp. bottom) row: difference to DDPM-avg (resp. DDPM); see Supplement for p-values). Arrows indicate whether high or low values are optimal. B) Pixel intensity profiles along the dashed yellow line (left) for all models (right). C) Top row: A representative low-intensity image (Raw) from the test dataset, the corresponding high-resolution version (ground truth, GT), and the results of our proposed method. Bottom row: results of the benchmark models. The scale bar indicates 2 µ m. D) Pixel-wise difference between the ground truth and the reconstruction for each model using the sample in C. Blue (red) values indicate a lower (higher) predicted pixel value.
• Figure 2: Conditioned DDPMs outperform all benchmark models in denoising STED images of live-cell mitochondria. A - D are as in Fig. \ref{['fig:results-tubulin']}, but for the live-cell mitochondria dataset. The scale bar in (C) indicates 1 µ m.
• Figure 3: Conditioned DDPMs outperform several previous methods in denoising confocal images of synapses in the mouse brain. (A), (B), (C), (E) are as in Fig. \ref{['fig:results-tubulin']} (A), (B), (C), (D), respectively, but for the synapse dataset. D) Shows the $xz$-view of a sample. The scale bar indicates 1 µ m.
• Figure 4: Conditioned DDPMs outperform previous methods in denoising confocal images of zebrafish embryos. (A), (B), (C), (D) are as in Fig. \ref{['fig:results-tubulin']}, but for the zebrafish dataset. The scale bar in (C) indicates 10 µ m.
• Figure S1: Averaging across samples improves the performance for most metrics for the DDPM. We repeatedly predict a denoised image using the same low-intensity conditioning input but different initial noise. We compute the mean image across different numbers of reconstructions. Average performance is shown in bold, and the translucent ban indicates the standard deviation.