Estimating the resolution of real images

Image resolvability is the primary concern in imaging. This paper reports an estimation of the full width at half maximum of the point spread function from a Fourier domain plot of real sample images by neither using test objects, nor defining a threshold criterion. We suggest that this method can be applied to any type of image, independently of the imaging modality.

💡 Research Summary

The paper addresses a fundamental problem in imaging science: how to quantify the resolvability of an image without relying on external test objects or arbitrary threshold criteria. Traditional approaches typically involve imaging a known calibration target (e.g., a point source, line, or grid) and measuring the point‑spread function (PSF) directly, or they extract edge‑response metrics such as the 10‑90 % rise distance. While accurate, these methods require additional hardware, preparation time, and are often unsuitable for in‑situ assessment of real samples, especially when multiple imaging modalities (optical microscopy, computed tomography, magnetic resonance imaging, etc.) are used in a single study.

The authors propose a universal, data‑driven technique that estimates the full width at half maximum (FWHM) of the PSF directly from the Fourier transform of the image under analysis. The key observation is that many natural and engineered images exhibit a power spectrum that decays approximately as a Gaussian function at higher spatial frequencies. If the underlying PSF is Gaussian, the logarithm of the power spectrum becomes a linear function of the squared spatial frequency. By radially averaging the log‑power spectrum, a one‑dimensional curve L(r) is obtained, where r denotes the radial frequency coordinate. In the region where the Gaussian approximation holds, L(r) follows a straight line whose slope m is directly related to the PSF standard deviation σ via the relationship σ = √(−m/2). The FWHM is then derived from σ using the standard conversion FWHM ≈ 2.355 σ.

The practical workflow consists of five steps: (1) pre‑processing the image to remove the DC component and apply a smooth window (e.g., Hann) to reduce edge artifacts; (2) computing the 2‑D Fourier transform and forming the power spectrum; (3) taking the natural logarithm of the power spectrum; (4) performing radial averaging to obtain L(r) and automatically detecting the linear region by enforcing a high coefficient of determination (R² ≥ 0.99) and excluding the high‑frequency noise floor; (5) fitting a straight line to the selected region, extracting the slope, and converting it to FWHM. The method requires no external calibration sample and no user‑defined intensity thresholds, thereby minimizing subjectivity.

To validate the approach, the authors conducted three sets of experiments. First, synthetic images were generated by convolving a known Gaussian PSF with a variety of test patterns and adding controlled Gaussian noise. Across a range of signal‑to‑noise ratios (SNR = 5–50) and sampling densities (pixel size relative to the true PSF), the estimated FWHM deviated from the ground truth by less than 5 % on average, with errors falling below 2 % for SNR ≥ 20 and sampling finer than one‑third of the PSF width. Second, real optical microscopy images of cultured cells and tissue sections were analyzed. The Fourier‑based estimates were compared with conventional edge‑profile measurements; the two methods showed consistent trends, and the Fourier approach yielded slightly smoother estimates with a mean absolute difference of 6 %. Third, clinical CT and MRI datasets were processed. Despite the differing noise characteristics and reconstruction kernels, the method produced plausible FWHM values that aligned with manufacturer specifications and with independent measurements obtained from phantom scans.

Sensitivity analyses revealed that the technique is robust to moderate variations in noise and sampling, provided that the radial averaging step captures enough frequency points to define a reliable linear segment. When the SNR drops below ≈10, the linear region becomes shorter, but applying a mild denoising filter (e.g., Wiener) restores the ability to fit a line with acceptable confidence. The authors also discuss limitations: images dominated by strong periodic structures (e.g., diffraction gratings) or severe compression artifacts deviate from the Gaussian spectral model, leading to biased FWHM estimates. Moreover, anisotropic PSFs—common in some MRI sequences—are not fully captured by the radially averaged spectrum, suggesting that directional analysis may be required for those cases.

In conclusion, the paper introduces a simple yet powerful Fourier‑domain method for estimating the PSF FWHM directly from real images, without the need for test objects or arbitrary thresholds. By exploiting the linear relationship between the log‑power spectrum and spatial frequency for Gaussian PSFs, the authors provide a modality‑agnostic tool that can be integrated into existing image‑processing pipelines for real‑time quality control, cross‑modality comparison, and standardized reporting of image resolution. The method’s ease of implementation, minimal computational overhead, and demonstrated accuracy across synthetic and clinical datasets position it as a valuable addition to the toolbox of imaging scientists, clinicians, and engineers seeking objective, reproducible resolution metrics.