Misc,
DAOSTORM: An algorithm for high-density super-resolution microscopy
(2011)
DOI: 10.1038/nmeth0411-279
Abstract
correspondence We first investigated the qualitative performance of each algorithm for images of Alexa Fluor 647-immunolabeled microtubules in fixed COS-7 cells. We recorded data at high imaging density using total internal reflection fluorescence microscopy and direct (d)STORM photoswitching conditions 5 (100 ms integration time, \~4,000 photons fluorophore-1 frame-1). We plotted localizations on raw images, illustrating the characteristic performance of each algorithm (Fig. 1a). SA1 only localized isolated molecules, which were fitted with small localiza-tion error. SA2 localized a larger fraction of the molecules but yielded large localization errors for overlapping molecules. DAOSTORM out-performed both sparse algorithms, identifying almost all molecules with small localization error. We quantified the performance of each algorithm by analyzing simulations of randomly distributed surface-immobilized fluorophores 6. We compared observed localizations to simulated positions, calculating the recall 5 and localization error at different imaging densities. Recall is the percentage of simulated fluorophores detected. Localization error is the root-mean-square distance between a localization and the simulated position. We also measured the precision 5 and redundancy (Supplementary Methods), which did not vary substantially. DAOSTORM substantially outperformed the sparse algorithms in simulations at high signal-to-noise ratio typical of STORM data (bright organic fluorophores, 5,000 photons molecule-1 frame-1 ; Fig. 1b-c). SA1 showed poor recall at high density, with imaging density at half-maximum recall, $\rho$ HM , of 1.2 molecule µm-2. However, SA1 yielded small localization errors even at high imaging density because most overlapping molecules were rejected. SA2 had better recall performance ($\rho$ HM = 3.4 molecules µm-2
