Structured Illumination Microscopy With Uncertainty-Guided Deep Learning

Abstract

Super-resolution microscopy enables the visualization of subcellular structures with unprecedented detail, significantly advancing life sciences. Among the various techniques available, structured illumination microscopy (SIM) provides an ideal balance of speed, resolution, and phototoxicity. Recent advancements in deep learning have further enhanced SIM capabilities, achieving improved imaging quality with higher signal-to-noise ratios and fewer measurements. However, the opaque nature of these deep learning models complicates the quantification of uncertainty in their outputs, which may lead to visually appealing but scientifically inaccurate results, particularly challenging for clinical diagnostics. In this paper, we introduce a two-step strategy that not only quantifies the uncertainty of deep learning models but also enhances super-resolution reconstruction. The first step implements a novel sparse-constrained loss function, incorporating Jeffrey's prior, to accurately predict uncertainty maps. These maps assess the confidence levels of the network's predictions and identify potential inaccuracies. In the second step, these predicted uncertainty maps serve as an attention mechanism, directing the neural network's focus towards areas of high uncertainty to improve the reconstruction of high-frequency details and textures. A series of simulations and experiments confirm that our method accurately quantifies uncertainty and improves high-resolution image reconstruction, increasing the peak signal-to-noise ratio by an average of 1.7 dB and structural similarity by 0.06, compared to traditional methods on mitochondrial and microtubule datasets. Our approach holds promise for advancing the application of deep learning-based super-resolution microscopy in clinical settings.