Fourier Shell Correlation, or FSC, is the standard method for estimating the resolution of three-dimensional density maps in cryo-electron microscopy. It works by splitting the raw data into two independent halves, reconstructing a 3D map from each, and then measuring how well those two maps agree at progressively finer levels of detail in Fourier space. Where agreement drops below a chosen threshold, the map is considered unreliable, and the spatial frequency at that point gets converted into a resolution number in angstroms. The concept is intuitive, but the practical details of how the split is done, which threshold to use, and how easily the metric can be gamed or inflated make FSC a richer and more contested subject than a single reported resolution might suggest.
How the Half-Map Split Works
The core idea behind FSC is correlation. You take your dataset of particle images, randomly divide it into two equally sized subsets, and independently reconstruct a 3D volume from each half. These are called “half-maps.” You then transform both volumes into Fourier space, where each point encodes information about a particular spatial frequency and direction. The Fourier coefficients are grouped into thin spherical shells by spatial frequency, and for each shell, you compute the normalized cross-correlation between the two half-maps. The result is a curve: FSC plotted against spatial frequency. At low spatial frequencies (corresponding to large, coarse features), the two half-maps almost always agree strongly, giving FSC values near 1.0. As spatial frequency increases toward finer detail, noise starts to dominate and correlation drops. The resolution of the map is wherever that drop crosses a pre-defined cutoff.
This approach directly estimates the spectral signal-to-noise ratio of the reconstruction. The FSC at any given shell is mathematically related to how much real signal exists at that frequency compared to noise, and this relationship has been studied for various classes of reconstruction algorithms.1PubMed. Three-dimensional spectral signal-to-noise ratio for a class of reconstruction algorithms The key requirement is that the two halves be statistically independent. If they share information, whether through shared processing steps or correlated noise, the FSC will overestimate how much genuine structural detail the map contains.
The 0.143 Threshold and Why It Is Debated
The most commonly used FSC cutoff in single-particle cryo-EM is 0.143. This value was originally derived to approximate the point at which a half-map FSC corresponds to an FSC of 0.5 for the full dataset, which in turn corresponds to a signal-to-noise ratio of about 1.0 at that shell. It became the default for reporting resolution, and virtually every structure deposited in the Electron Microscopy Data Bank uses it.
Not everyone agrees that 0.143 is the right number. A detailed critique by Marin van Heel and colleagues argues that the derivation of the 0.143 threshold left out important noise cross-terms, making it systematically too generous. Their analysis contends that the threshold is “itself incorrect, even in an asymptotic sense, because the signal-noise cross-terms have explicitly been left out during its derivation.”2bioRxiv. Reassessing the Revolution’s Resolutions According to this view, a more conservative cutoff, such as 0.5, would better reflect the true resolution of many reconstructions.
In practice, the community has largely stuck with 0.143 for standardized reporting, though reviewers and readers have become more aware that it can be optimistic. The threshold is just one piece of a larger picture, and experienced structural biologists rarely rely on a single FSC number without checking other indicators of map quality. Whether 0.143 or 0.5 is used, the fundamental limitation is the same: a single number summarizes a whole curve, and that summary inevitably loses information about how resolution varies across the map and in different directions.
Gold-Standard FSC and Overfitting
One of the most insidious problems in cryo-EM reconstruction is overfitting, where the refinement algorithm starts fitting noise rather than genuine structural features. Because iterative refinement uses the data itself to improve the model, high-resolution noise correlations can creep in, producing maps that look sharp but contain false detail. The FSC curve for an overfitted map can look deceptively good.
The gold-standard FSC approach was proposed to combat this. Instead of splitting the data only at the end for the resolution check, the two halves are kept completely separate from the very beginning of processing. Each half-dataset is refined independently against its own evolving reference. This means neither half ever “sees” the other’s noise during refinement, so any correlation between the two final half-maps reflects real signal rather than shared fitting artifacts.3PubMed Central. Prevention of overfitting in cryo-EM structure determination Gold-standard FSC has become the default in major processing software and is expected by journals and database depositions.
A complementary strategy is high-resolution noise substitution. Here, the Fourier components of the particle images beyond a chosen resolution are replaced either with background noise from the micrograph or by randomizing the phases. The refinement is then run again with this degraded data, and the resulting FSC curve is compared with the original. Any residual correlation at frequencies where real signal was removed is direct evidence of overfitting. A correction formula allows researchers to calculate an unbiased FSC even when overfitting is present.4Ultramicroscopy. High-resolution noise substitution to measure overfitting and validate resolution in 3D structure determination by single particle electron cryomicroscopy
How Masks Inflate Resolution Estimates
Before computing the FSC, researchers typically apply a soft mask around the molecule of interest to exclude surrounding solvent noise. This step seems harmless and often genuinely improves the measurement by preventing solvent noise from diluting the signal. But masking introduces a subtle and serious artifact: multiplying two volumes by the same mask function creates artificial correlations between them in Fourier space.
The mechanism is a consequence of the convolution theorem. Multiplying a volume by a mask in real space is equivalent to convolving its Fourier transform with the Fourier transform of the mask. This convolution smears Fourier coefficients across neighboring frequencies, which introduces correlations between shells that would otherwise be independent. Even if the two half-maps were pure independent noise, applying the same mask would produce FSC values significantly above zero, falsely suggesting the presence of signal.5IUCrJ. Reliable cryo-EM resolution estimation with modified Fourier shell correlation
Several correction strategies exist. The most common is the phase-randomization test introduced by Peter Rosenthal and Richard Henderson, where phases are randomized beyond a certain resolution, the masked FSC is recalculated, and the spurious mask-induced correlation is subtracted from the original curve. A more recent approach, the modified FSC, simply changes the order of computational steps to eliminate mask-induced correlations altogether, rather than trying to correct them after the fact. This modified procedure has been shown to produce reliable, artifact-free results for both single-particle and helical structures.6PubMed Central. Reliable cryo-EM resolution estimation with modified Fourier shell correlation
Local Resolution Across the Map
A global FSC curve assigns a single resolution value to the entire 3D map. That is a useful summary, but it hides the fact that resolution often varies dramatically from one part of the structure to another. A rigid core domain might be resolved at 2.5 angstroms while a flexible loop or a peripheral subunit reaches only 5 or 6 angstroms. Reporting only the global number misrepresents both regions.
Local resolution methods address this by computing FSC-like comparisons in small, overlapping windows across the volume rather than over the whole map at once. The result is a three-dimensional map of resolution values that can be color-coded and overlaid on the density. This has changed how cryo-EM maps are interpreted, providing information about heterogeneity, flexibility, and angular assignment errors, and guiding decisions about where atomic models can be built with confidence.7PubMed Central. Local resolution estimates of cryoEM reconstructions
The global FSC measurement falls short of describing these local differences, and filtering an entire map to a single resolution cutoff can blur well-resolved regions while leaving poorly resolved regions looking noisier than they need to be.8PubMed Central. Model-based local density sharpening of cryo-EM maps Local resolution estimates are now routinely reported alongside global FSC values in structure depositions and publications.
Directional Resolution and Preferred Orientation
Resolution can also vary with direction, not just with location in the map. This happens when particle images are not evenly distributed across all viewing angles. If a molecule has a strong tendency to land on the grid in one orientation, the reconstruction will have excellent resolution along certain directions and poor resolution along others. The global FSC averages over all directions on each spherical shell, so it can report a respectable number even when resolution is severely anisotropic.
Conical FSC methods detect this problem by restricting the correlation calculation to cone-shaped subsets of Fourier space, each pointing in a different direction. Instead of averaging over an entire shell, the FSC is computed separately for each cone orientation, revealing which directions are well-determined and which are not.9bioRxiv. cFAR and Relative Signal: Diagnosing Preferred Orientation in Single-Particle Cryo-EM A metric called the conical FSC area ratio, or cFAR, summarizes this as the ratio of the worst direction’s performance to the best direction’s. A value of 1 means perfectly isotropic resolution; as anisotropy worsens, the ratio drops toward 0.
Directional resolution assessment has been particularly useful in electron tomography, where the missing wedge problem inherently limits information along the beam direction. Comparing single-axis and dual-axis tilt schemes with conical FSC has shown that dual-axis acquisition produces more isotropic resolution, with high-resolution peaks along the tilt axis disappearing.10Journal of Structural Biology. Conical Fourier shell correlation applied to electron tomograms
Map-to-Model FSC
The half-map FSC tells you how well the two independent reconstructions agree with each other, but it does not directly tell you whether the atomic model you build into the map is correct. A separate validation tool, map-to-model FSC, addresses this by comparing the experimental cryo-EM map against a simulated map generated from the fitted atomic coordinates.
The simulated map is calculated from the model at a resolution comparable to or finer than the experimental map. Because the model-derived map contains no noise, the map-to-model FSC should cross the 0.5 threshold at roughly the same resolution where the half-map FSC crosses 0.143. If the model FSC drops off significantly earlier, the model may be overfitted or built into regions where the map does not actually support it. If it extends further, the model may be too smooth.11Acta Crystallographica Section D Structural Biology. Validation, analysis and annotation of cryo-EM structures Comparing the two curves side by side has become a standard part of cryo-EM structure validation, and reviewers routinely expect to see both plots.
FSC-Guided Map Improvement
Beyond measuring resolution, the FSC curve itself can be used as a tool for improving map quality. Cryo-EM maps straight out of reconstruction software tend to have suppressed high-frequency amplitudes due to the effects of the contrast transfer function, radiation damage, and other factors. Sharpening, usually by applying a negative B-factor, boosts these high frequencies to restore contrast. The FSC curve helps determine how much boosting is appropriate and where to stop before amplifying pure noise.
More sophisticated approaches use local sharpening guided by local resolution estimates, or use atomic models to calibrate the expected amplitude profile and then scale the map locally to match. These model-based methods can optimize contrast across the entire map rather than applying a single uniform correction.8PubMed Central. Model-based local density sharpening of cryo-EM maps
Deep-learning post-processing tools have also emerged that learn the relationship between noisy input maps and their ideal counterparts. One such tool, DeepEMhancer, improved the similarity between processed maps and reference maps derived from atomic models by a median of roughly 0.6 angstroms in resolution, or about 14% in the frequency domain.12PubMed Central. DeepEMhancer: a deep learning solution for cryo-EM volume post-processing The FSC curve remains central to evaluating whether any post-processing step has genuinely improved the map or merely made it look different.
A related development is self Fourier shell correlation, which estimates the signal-to-noise ratio from a single volume rather than requiring a pair of half-maps. The self-FSC profile can serve as a Wiener-like filter, weighting each Fourier shell according to its estimated signal content to denoise a reconstruction.13PubMed Central. Self Fourier shell correlation: properties and application to cryo-ET This is especially useful in cryo-electron tomography, where splitting limited tilt-series data into two independent halves can compromise both reconstructions.
Fourier Ring Correlation in Optical Microscopy
The two-dimensional sibling of FSC is Fourier ring correlation, or FRC. The idea is identical: split the data, compute two independent images, and measure their correlation in rings of increasing spatial frequency. FRC has found a natural home in super-resolution fluorescence microscopy, where the question of how much genuine structural detail an image contains is just as pressing as in cryo-EM but historically lacked a standardized quantitative answer.
In localization microscopy techniques such as STORM and PALM, FRC can be computed by randomly dividing the detected localizations into two halves and reconstructing an image from each. The correlation between the two images as a function of spatial frequency gives an objective, data-driven resolution estimate that accounts for both localization precision and labeling density. This has been validated on images of cellular structures like tubulin and actin filaments in both two and three dimensions.14Nature Methods. Measuring image resolution in optical nanoscopy Freely available software implementing FRC-based resolution measurement has helped it become a laboratory-consistent standard across different super-resolution methods.15PubMed. Fourier ring correlation as a resolution criterion for super-resolution microscopy
The adoption of FRC in light microscopy underscores how broadly useful the correlation-based framework is. Whether you are looking at a single protein complex with electrons or a whole cell with fluorescent labels, the underlying logic is the same: compare independent measurements, find where they stop agreeing, and call that your resolution. The specifics differ, particularly around how the data split is performed and what threshold is appropriate, but the conceptual foundation transfers directly from cryo-EM’s three-dimensional shells to optical microscopy’s two-dimensional rings.