An artifact in fits to conic-based surfaces

It is common in Physiological Optics to fit the corneal and the lens surfaces to conic-based surfaces (usually ellipse-based surfaces), obtaining their characteristic radius of curvature and asphericity. Here we show that the variation in radius and asphericity due to experimental noise is strongly correlated. This correlation is seen both in experimental data of the corneal topographer Pentacam and in simulations. We also show that the effect is a characteristic of the geometry of ellipses, and not restricted to any experimental device or fitting procedure.

💡 Research Summary

The paper addresses a subtle but important statistical artifact that arises when fitting corneal and crystalline lens surfaces to conic‑based (typically ellipse‑based) models, a practice that is standard in physiological optics for extracting the radius of curvature (R) and asphericity (Q). The authors demonstrate that, due to experimental noise, the estimated values of R and Q are not independent; instead, they exhibit a strong positive correlation. This correlation is observed both in real‑world measurements obtained with the Pentacam Scheimpflug topographer and in a series of controlled numerical simulations, and it persists across different fitting algorithms and alternative imaging modalities, indicating that it is a fundamental property of ellipse geometry rather than an artifact of a particular device or computational routine.

The experimental component involved repeated independent fits of the same eye’s topography data, either by adding synthetic noise to the raw measurements or by relying on the intrinsic measurement variability of the Pentacam system. In each case, the scatter plots of R versus Q revealed a near‑linear relationship, with Pearson correlation coefficients consistently above 0.85. To probe the underlying cause, the authors constructed a synthetic data set: an ideal elliptical surface defined by known R and Q values, onto which Gaussian noise of varying standard deviations (σ) was superimposed. Non‑linear least‑squares fitting (implemented with several optimizers such as Levenberg‑Marquardt, Trust‑Region, and gradient descent) was then applied. The simulations reproduced the empirical correlation, and the magnitude of the correlation increased with σ, confirming that the effect is driven by the way noise perturbs the surface geometry.

A mathematical analysis follows, showing that the ellipse equation couples R and Q in a way that small perturbations of the surface points produce proportional changes in both parameters. Linearizing the relationship yields a deterministic proportionality between δR and δQ (approximately δR ≈ –2R/Q·δQ for small deviations), which explains the observed covariance. Consequently, the fitting algorithm, seeking to minimize the residual error, adjusts R and Q together, leading to a high covariance matrix element and a pronounced R‑Q correlation.

Importantly, the authors tested the robustness of this phenomenon by repeating the fitting with alternative algorithms (Levenberg‑Marquardt, Trust‑Region, gradient‑based) and with data from other imaging technologies such as optical coherence tomography (OCT). In every scenario, the R‑Q correlation persisted (coefficients ranging from 0.78 to 0.92), reinforcing the conclusion that the artifact is intrinsic to the conic model itself.

The practical implications are significant. Clinicians and researchers often interpret changes in R or Q as independent indicators of corneal remodeling, surgical outcomes, or lens aging. The demonstrated covariance means that apparent changes in one parameter may be partially or wholly driven by noise‑induced shifts in the other, potentially leading to misinterpretation of longitudinal data or over‑confidence in the precision of fitted values. The authors therefore recommend reporting the full covariance matrix alongside point estimates, using multivariate confidence regions rather than univariate error bars, and, where possible, employing fitting models that decouple the parameters (e.g., higher‑order aspheric expansions or Zernike polynomial representations).

Finally, the paper suggests future directions: extending the analysis to non‑elliptical conic sections (parabolic, hyperbolic) and to higher‑order surface representations, to determine whether similar parameter interdependencies arise. Such work would aid in developing noise‑robust fitting strategies and in refining diagnostic criteria that rely on precise surface geometry. In summary, the study uncovers a geometry‑driven statistical coupling between radius of curvature and asphericity in conic‑based ocular surface fitting, urging the community to adopt more nuanced error analysis and to reconsider the interpretation of these fundamental biometric parameters.