betaFeedback

WorksPhysics and Astronomy

Three-dimensional stochastic retinal models for synthetic eye-models

Abstract

Purpose: Current eye models often assume a fixed spherical retinal curvature of 12 mm, which does not reflect the anatomical variability observed in vivo. This study proposes an extension incorporating the stochastic relationship between axial length and three-dimensional (3D) retinal curvature, enabling more realistic 3D eye models. Methods: Axial length and 3D retinal curvature data from 94 subjects, split into five subsets for cross-validation, were used to construct two probabilistic models: a Single Gaussian Model (SGM) and two-component Multi Gaussian Mixture Model (MGMM). Both were converted into conditional forms to estimate retinal curvature for a specific axial length, and evaluated against measured data distributions. Results: All models produced radii of curvature comparable to measured data (median difference <0.2 mm). Non-conditionally, the MGMM described the distribution slightly better than SGM, with higher density overlap (>94% vs. >86%), but similar variance (worst f-test (p-value) 1.05 (0.71) vs. 1.05 (0.73)). Conditioned on axial length, both models performed similarly with overlap exceeding 93% and the worst f-test 1.02 (0.84). Given its simplicity and robustness, the conditional SGM was selected, achieving similar validation results: median difference (<0.10 mm), overlap (>92%), and a non-significant worst f-test (1.02 (0.85)). Conclusions: The conditional SGM enables accurate generation of anatomically realistic 3D retinal curvatures based on axial length, supporting improved stochastic eye modeling for visual optics and population-based simulations of healthy eyes. Translational relevance: Replacing the fixed 12 mm retina assumption with a population-based stochastic model improves anatomical accuracy of eye models used in optical design and treatment planning.

Corresponding Author:

Nadine P. van Dam

Department of Ophthalmology

Leiden University Medical Center

Albinusdreef 2

Leiden, 2333 ZA, The Netherlands

[email protected]

Funding: Dutch Research Council (NWO), Grand Number: OTP-CAPTOR, 21360.

Commercial Relationship Disclosure: Nadine P. van Dam, Trident S.r.l., and MedCom GmbH; Luc van Vught, Current employee of Cassini Technologies B.V.; Jos J. Rozema, Consultant for Azalea Vision and Morrow Vision; Steven J. M. Habraken, Trident S.r.l., and MedCom GmbH; Jan-Willem M. Beenakker, Philips Healthcare, Trident S.r.l., and MedCom GmbH.

Introduction

An accurate model of the ocular geometry is essential for studying the optical properties of the eye and developing effective therapeutic strategies. Over the years, several schematic eye models have been proposed, including the Gullstrand [1], Le Grand [2], Emsley [3], and Navarro [4] models [5], which rely on a single set of parameters to describe ocular surfaces. Each of these models was designed to accurately capture a specific property of the eye, such as higher-order aberrations or peripheral refraction. These models have served as a foundation for understanding vision complaints such as negative dysphotopsia and for guiding improved treatment strategies, including intra-ocular lenses that maintain peripheral vision quality [6, 7, 8, 9, 10, 11].

In contrast to these schematic models, real eyes exhibit a wide variety of anatomical variations [12, 13, 14, 15]. The statistical relationships between key biometric parameters, such as axial length, corneal curvature, and anterior chamber depth, have been extensively documented [16, 17, 18, 19, 20] and form the basis of nomograms that guide clinical decision-making, such as selecting an intraocular lens for cataract surgery. Rozema et al. introduced SyntEyes [21, 22], a stochastic eye model that unlike conventional eye models, considers the statistical distribution of biometric parameters and their correlations. Instead of specifying a single eye model, a stochastic eye model generates a virtual cohort of eyes that realistically capture the variability in ocular biometry within the normal population. These generated eyes have been used in vision science to simulate the correction of refractive errors in healthy eyes [23] or eye tracking [24] with different biometric parameters, without the need for real clinical data. Additionally, SyntEyes have also been designed for specific clinical conditions such as keratoconus [25], and were used for similar corrective simulations [26].

Although the SyntEyes model captures most ocular biometry parameters relevant for central vision, they assume a spherical retina with a fixed 12 mm radius of curvature, similar to other eye models [27, 28]. However, several studies have shown that the retinal surface is more accurately represented by an ellipsoidal shape and that its radii differ significantly between individuals and from the generic value of 12 mm, with standard deviations up to 0.7 mm [9, 29, 30, 31, 32]. These variations in retinal geometry influence clinically relevant functional characteristics, including peripheral refraction [32], prediction of myopia progression [33], and the quantification of fundus imaging [34]. Consequently, the applicability of SyntEyes for assessment beyond central vision is limited.

It is hypothesized that the Multivariate Gaussian Mixture Model (MGMM), which underpins the SyntEyes framework, can also be applied to accurately describe the statistical relationship between axial length and the retinal radii, as measured by high-resolution Magnetic Resonance Imaging (MRI). Leveraging this approach would enable the extension of SyntEyes to incorporate three-dimensional retinal shape information, providing a more anatomically accurate representation of the posterior segment of the eye.

Methods

Participants

In this study, 94 individuals with no relevant ophthalmic history (e.g. no pathology or condition), except for cataract surgery in a subset of participants were retrospectively included. Ocular biometry (Lenstar LS900; Haag-Streit AG, Koeniz Switzerland), with i.e. the axial length defined as the distance from the cornea to inner limiting membrane, and high-resolution MRI scans were prospectively acquired in the context of other clinical studies, each of which adhered to the Declaration of Helsinki and received approval by the Medical Ethics Committee Leiden The Hague Delft [6, 9, 10, 29]. All the participants provided written informed consent before their inclusion into the studies.

The retinal curvature parameters were derived from high-resolution MRI, as previously described [9, 29]. All MRIs were acquired with a Philips Achieva 7T whole-body magnet (Philips Healthcare, Best, The Netherlands) equipped with an in-house developed eye receive coil [9, 29, 35]. Retinal segmentation was carried out automatically using subdivision surfaces in MeVisLab (MeVis Medical Solutions AG, Bremen, Germany) [9, 29]. Subsequently, a three-dimensional ellipsoid was fitted to the posterior 240 degrees of the retinal contour, following the method proposed by van Vught et al. [29], Figure 1.

Figure 1. Illustration of the included parameters shown in the inferior-superior plane of a T1-weighted MRI for a phakic eye. The green and pink circles correspond to the 12 mm spherical retinal curvature as reported in literature and the retinal curvature reported by van Vught et al. [29], respectively. The dashed part shows 240 degrees of the retina used to fit the model. The sky blue continuous line represents the axial length (24.31 mm), defined as length from cornea to inner limiting membrane, measured with the Lenstar, which is in agreement with the MRI [29]. The yellow and light blue dashed lines provide the RNTR_{\mathrm{NT}} (12.24 mm) and RPAR_{\mathrm{PA}} (10.82 mm) radii in the nasal-temporal and posterior-anterior directions, respectively. The orange circle represents the location of the inferior-superior radius (RISR_{\mathrm{IS}}=11.55 mm).

Models

The large standard deviation in the retinal curvature reported in literature, suggests that a linear regression model would not sufficiently describe the distribution in the healthy population. Therefore, a more complex Gaussian Model was fitted on the data. Since it is unknown whether the relationship between axial length and retinal curvature necessitates two Gaussian components, both a Single Gaussian Model (SGM) [21] and a two-component MMGM [22] were evaluated. The SGM and MGMM are implemented in Python (v 3.13) using the scipy (v 1.15.2) and scikit-learn (v 1.6.1), respectively.

A conditional version of the retinal curvature model, based on a given axial length, was derived such that the retinal model can integrated into the existing SyntEyes model [22] (see Supplement 1 for the mathematical derivation). Using this formalism, 3D SyntEyes can be constructed by first generating the conventional SyntEyes parameters and then calculating the matching retinal radii for each eye, using the conditional retinal GM.

In this paper, both the non-conditional models, which do not take the axial length as input, and the conditional models, which take the axial length into consideration, were evaluated.

Statistics

Before fitting the models on the data, it was determined how many eyes in the measured dataset approximately followed the 12 mm sphere (Euclidian distance between sphere and ellipsoid ≤0.5\le0.5mm), to underscore the importance of creating a more realistic model. Additionally, the effect of including pseudophakic eyes was investigated with a Bonferroni corrected Mann-Whitney-U test.

The performance of each model was tested with five-split cross-validation. Given the limited sample size, this approach was chosen to avoid overfitting and improve the generalizability of the resulting models [36]. Data splitting was performed randomly with the sole constraint of an equal distribution of pseudophakic eyes, resulting in four splits of 75-19 and one split 76-18 of training-validation data points. A Bonferroni corrected Mann-Whitney-U test was performed across the different splits to access whether the resulting training and validation sets formed similar distributions.

The trained models for the different splits were then used to generate datasets to compare the models with the measured data. For each eye included from the measured dataset in the split 100 eyes were generated with the corresponding model. This approach was chosen such that the generated dataset clearly represents the models’ distribution.

Model performance was quantified with three measures. First, the difference in median was calculating by subtracting the model’s median from the measured median. After that, the model’s variance was compared to the measured variance using an f-test of equality of variance. Lastly, the overlap in Probability Density Functions (PDFs) between the data generated by the models and the training and validation data. A Gaussian kernel was used, with Scott’s rule adopted to determine the appropriate bandwidth [37]. As the small dataset size limits the stability of the metrics, particularly for the validation datasets, the metrics are not determined for each split separately but on the entire generated cohort.

Lastly, the final model was assessed using the median absolute distance in mm between the radius Cumulative Density Functions (CDFs) of the measured and generated validation data, within the 0.05-0.95 cumulative range, excluding tail influences.

Retinal model application

The resulting average SGM model was integrated in the open-source ophthalmic raytracing platform Visisipy [38] to investigate the effect of the retinal model on the resulting relative peripheral refraction. The Navarro wide-angle eye [27] was taken as the basis and the relative refraction at 40∘40^\circ was calculated for 1000 conditional SGM eyes with the same axial length (23.92 mm) but varying retinal radii.

Results

Population cohort

The total cohort consisted of 94 individuals with a wide range in age (19–83 years) and axial length (20.5–28.4 mm). About two-thirds of the subjects were pseudophakic, and other descriptors of the cohort, such as the retinal radii, can be found in Table 1. In this dataset, only 3/94 (3%) of the retinal ellipsoids could be approximated as a sphere with a 12 mm radius. The Mann-Whitney-U test demonstrated no significant differences for any of the study parameters between the phakic and pseudophakic subjects (p>0.05/4=0.0125p > 0.05/4 = 0.0125, Bonferroni correction).

Measured dataMedian (min-max)
Participants94
Sex (female)65
IOL61
Eye (left)65
Age [years]63.2 (18.8-82.6)
ALAL [mm]23.6 (20.5-28.4)
RNTR_{NT} [mm]11.9 (10.4-14.6)
RISR_{IS} [mm]11.6 (10.3-13.0)
RPAR_{PA} [mm]10.4 (9.2-12.4)

Table 1. Table 1 Characteristics of the dataset. In this table, IOL represents the number of participants with an intraocular lens implantation. The ALAL, RNTR_{NT}, RISR_{IS}, and RPAR_{PA} represent the axial length and the retinal radii in the nasal-temporal, inferior-superior, and posterior-anterior direction, respectively.

The distribution of axial length and the posterior-anterior radius resembles a two-component Gaussian distribution, while the nasal-temporal and inferior-superior radii more closely resemble a single-component Gaussian distribution (Figure 2). Although all four studied parameters are correlated, the strongest correlation was observed between axial length and the posterior-anterior radius (rpearson=0.862r_{\mathrm{pearson}} = 0.862).

Measured data compared to the non-conditional models

Figure 2. The diagonal gives the distribution of the measured, the non-conditional SGM-generated, and MGMM-generated data for each biometric parameter in grey, blue, and orange, respectively. Furthermore, in the radii plots, green vertical lines are shown at 12±0.512 \pm0.5 mm representing, the neighborhood of the 12 mm radii from literature. Additionally, the median distance (Δ\Delta), f-test (f) and overlap (o) are shown. The other plots provide insight into the relationship between the different biometric parameters, with the grey dots representing the measured data. The blue and orange lines are contour plots of the SGM and MGMM across all 5 splits combined, respectively. The outermost contour includes 95% of the generated data, and the inner contours represent 75%, 50%, and 25%. In the non-diagonal images, the Pearson correlation and corresponding p-value are provided.

The five training and validation datasets showed similar distributions as the Mann-Whitney-U test was not significant for all parameters and between every split (p>0.05/40=0.00125p > 0.05/40 = 0.00125, Bonferroni correction). Further details of the splits are provided in Supplement 3.

Non-conditional models on the training dataset

The non-conditional MGMM performs better for the axial length and posterior-anterior radii than the SGM as seen from both the median difference (ALAL -0.04 mm vs. -0.27 mm and RPAR_{PA} -0.05 mm vs. -0.17 mm) and overlap (ALAL 0.94 vs. 0.89 and RPAR_{PA} 0.94 vs. 0.86) while the f-test was similar for both models (Table 2). For the nasal-temporal and inferior-superior radii, the performance differences between the models was smaller and both models show a small median difference and overlap values over 91%. The slightly better performance of the MGMM was also confirmed with visual inspection (Figure 2). Furthermore, the distribution of the SGM stayed consistent across all splits, whereas the MGMM components differed between splits (Supplement 3), suggesting overfitting. If the splits would be assessed separately, similar conclusions would have been drawn (Supplement 4).

Δmedian\Delta_{\mathrm{median}} [mm]F-test (p-value)Overlap [-]Δmedian\Delta_{\mathrm{median}} [mm]F-test (p-value)Overlap [-]
ParametersSGMMGMM
TrainingNon-conditional model
TrainingALAL-0.271.05 (0.71)0.89-0.041.05 (0.73)0.94
TrainingRNTR_{NT}-0.020.99 (0.99)0.910.061.01 (0.92)0.94
TrainingRISR_{IS}-0.031.01 (0.89)0.94-0.021.00 (0.96)0.95
TrainingRPAR_{PA}-0.171.01 (0.93)0.86-0.051.01 (0.89)0.94
TrainingConditional model
TrainingRNTR_{NT}0.011.01 (0.90)0.930.051.00 (0.94)0.94
TrainingRISR_{IS}-0.011.01 (0.90)0.96-0.021.02 (0.84)0.95
TrainingRPAR_{PA}-0.091.02 (0.85)0.93-0.041.01 (0.93)0.94
ValidationConditional model
ValidationRNTR_{NT}0.011.02 (0.85)0.92---
ValidationRISR_{IS}-0.011.01 (0.90)0.97---
ValidationRPAR_{PA}-0.101.02 (0.87)0.93---

Table 2. The overlap between the PDFs, the f-test of the measured data and eyes generated by the different models for the biometric parameters. This includes the axial length (ALAL) and the retinal radii in the nasal-temporal (RNTR_{NT}), inferior-superior (RISR_{IS}), and posterior-anterior (RPAR_{PA}) direction. There was no significant difference in the variance (p>0.05p > 0.05) between the measured and generated data.

Conditional models on the training dataset

A comparable evaluation as on the non-conditional training data was performed on the conditional models, in which a set of retinal radii was determined for each measured axial length and compared with the measured radii. The SGM model performed clearly better than the non-conditional counterpart with median difference for the posterior-anterior reduced to -0.09 mm and overlap of 93%. For the other radii the median difference was still small (0.01 mm and -0.01 mm for RNTR_{NT} and RISR_{IS}, respectively) and the overlap had increased (0.93 and 0.96, respectively). For all radii, the f-test showed that there was no significant difference in variance between generated and measured data (p>0.05p > 0.05). The performance of the MGMM stayed similar as the non-conditional counterpart. Furthermore, the performance between the SGM and MGMM was comparable which was also observed in the splits separately (Supplement 4).

Conditional model on the validation dataset

Given the similar performance of both models and the robustness of the less complex SGM, the validation was only performed on the SGM. Figure 3 provides the distributions, combined from all five splits, of the measured data and the conditional SGM model. The performance is comparable to the training data with median differences of 0.01, -0.01, and -0.10 mm for nasal-temporal, inferior-superior, and posterior-anterior, respectively. The f-test was still not significant (p>0.05p > 0.05). Additionally, the overlap exceeded 92% for all three radii, suggesting that the models indeed capture the data correctly without overfitting. In the validation dataset, the performance for each split separately is unstable due to the small amount of datapoint in the dataset (Supplement 4). Furthermore, the median absolute distance between the CDF for the left-right, inferior-superior, and posterior-anterior direction are 0.06, 0.02, and 0.07 mm, respectively.

Distribution of conditional models on the validation data

Figure 3. Comparison between the distributions of RNTR_{NT}, RISR_{IS}, and RPAR_{PA} for the conditional SGM model when the AL from the validation data was used as input. For the radii, the distributions of the measured validation data and the SGM are shown in gray and purple, respectively. In the corner, the median difference (Δ\Delta), f-test, and overlap provided.

Retinal model applied

The 3D SyntEyes were successfully implemented in Visisipy and 1000 eyes were generated based on the Navarro schematic eye, but with retinal radii based on the average conditional SGM (Supplement 2). For its axial length of 23.92 mm the resulting standard deviation in retinal radii were 0.43, 0.38, and 0.35 in the nasal-temporal, inferior-superior, and posterior-anterior direction, respectively. When for these 1000 eyes the wavefront aberrations at 40° in the inferior-superior and posterior-anterior plane was calculated, this resulted in a median peripheral refraction of 0.99 D, with an inter-quartile range of 1.38 D (see Figure 4). The 12 mm spherical retinal curvature resulted in a peripheral refraction of 1.46 D.

Raytracing and relative peripheral refraction

Figure 4. Raytracing at 40° was performed on 1000 eyes based on Navarro wide-angle model with the conditional SGM. On the left, three different retina curvatures are shown which shows that the ray is affected by the shape at the peripheral retina. On the right, the PDF of the relative peripheral refraction for the 1000 eyes is provided with the green vertical line representing the results for the 12 mm spherical retinal curvature used in literature.

Discussion

In this study, a stochastic model was established for the relationship between axial length and three-dimensional retinal radii in healthy eyes. Both an SGM and an MGMM were evaluated as non-conditional and conditional (on the axial length) stochastic models. Consistent with previous work, the axial length distribution in our cohort exhibited a bi-Gaussian character, reflecting the presence of a myopic subgroup [39]. Despite this, a single Gaussian model proved sufficient to derive statistically representative retinal radii when conditioned on axial length. When the conditional SGM is combined with the previously published SyntEyes model [21, 22], this approach enables the generation of three-dimensional eye models that more faithfully represent anatomical variation in the posterior segment of the healthy population.

The proposed conditional model demonstrated a close agreement with the measured data. Across all three retinal radii of curvature, the overlap between generated and measured distributions exceeded 92%, indicating that the model captures not only the average eye but also the clinically relevant spread of retinal geometries. Importantly, the absolute median CDF distance for each radius of curvature was at most 0.07 mm. When this value was compared to the standard deviation observed with the Navarro eye (minimal 0.35 mm), the CDF distance is small. Furthermore, the ray-tracing simulations showed that for an average eye the variation in retinal radii alone, results in a IQR of 1.38 D in peripheral refraction at 40° supporting the clinical relevance of a stochastic retinal model.

A limitation of this study is the relatively small dataset of 94 subjects, which precluded the use of an independent validation cohort and formal goodness-of-fit testing with metrics such as the two one-sided t-test employed in earlier SyntEyes studies [21, 22]. Although the dataset spans a wide range of axial lengths and retinal curvatures, a larger sample size would allow independent validation and the exploration of more flexible statistical representations, including the t-distribution [40] the skewed-normal distribution [41], or non-parametric mixture components [42]. Such models may better capture the underlying distribution, particularly for myopic eyes, where axial length and, by extension, retinal curvature is expected to be more complex than a single Gaussian. Nevertheless, the strong performance of the present conditional Gaussian model suggests that it already captures the measured variation in the data without evidence of clinically relevant overfitting and so outperforms the 12 mm spherical retina often assumed. Furthermore, the cross-validation did allow to investigate the robustness of the model despite the lack of independency and showed that the MGMM model was less robust. This observation can be explained by the large increase in fitting parameters for the MGMM versus SGM for which more data is needed to create a stable result.

It is important to note that the proposed stochastic retinal model incorporates only axial length as an input. Although other biometric parameters, such as anterior chamber depth, corneal curvature, refractive error, and off-axis refraction, may also be associated with retinal curvature. These relationships are expected to be weaker than that of the axial length. This is further supported by the strong agreement between the generated and measured retinal radii.

The proposed three-dimensional statistical eye model enables several clinically and technologically relevant applications. In refractive and cataract surgery, the extended SyntEyes framework allows population-based optical simulations that incorporate not only variability in corneal topography and axial length but also the associated distribution of retinal curvatures. This is particularly relevant for the design and evaluation of intraocular lenses aimed at preserving or optimizing peripheral vision [43], for understanding perceived visual phenomena such as (negative) dysphotopsia [8, 9], and for estimating the error in optical imaging modalities designed with the 12 mm assumption [34]. By accounting for realistic retinal geometry, the present model enables more comprehensive virtual cohorts for optical design studies, reducing reliance on simplified assumptions such as a fixed 12 mm retinal radius. To facilitate broader adoption in visual optics research, an online demo is created at demo.mreye.nl/synteyes, where 3D SyntEyes can be created with a retinal curvature determined using the axial length from the SyntEyes variables. The underlying code is available at https://github.com/MREYE-LUMC/3dsynteyes. These 3D SyntEyes have also been integrated into Visisipy, an open source raytracing platform specifically developed for visual optics simulations [38].

Beyond refractive surgery, the model supports applications in biomechanical modeling and ocular radiotherapy. The availability of realistic, population based posterior segment geometry provides a physiologically meaningful starting point for simulations of ocular growth and mechanical remodeling, for example in conditions such as keratoconus [25, 44], myopia progression [31, 33] and understanding the principles of crystalline lens wobbling [45]. In radiotherapy, statistical eye models may aid the development of more data-driven safety margins [46], particularly where uncertainties arise from mapping two-dimensional fundus images onto three-dimensional anatomical models used for treatment planning [34]. Finally, the availability of scalable, population-based 3D eye models facilitates generating large, anatomically plausible synthetic datasets for training AI algorithms in visual optics and ophthalmic imaging [47, 48, 49].

Conclusions

In conclusion, the retinal curvature can be modeled accurately with a SGM conditioned on the axial length. This opens the path for more representative eye simulations across the peripheral retina.

Supplement 1

The Gaussian Mixture models can be made conditional by determining a new mean, covariance, and weights for the unknown parameters with the following equations

μu,k∣xm=μu,k+Σum,kΣmm,k−1(xm−μm,k),(1)\boldsymbol{\mu}_{u,k \mid x_m} = \boldsymbol{\mu}_{u,k} + \boldsymbol{\Sigma}_{um,k}\boldsymbol{\Sigma}_{mm,k}^{-1}(x_m-\boldsymbol{\mu}_{m,k}), \tag*{(1)}
Σuu,k∣xm=Σuu,k−Σum,kΣmm,k−1Σum,kT,(2)\boldsymbol{\Sigma}_{uu,k \mid x_m} = \boldsymbol{\Sigma}_{uu,k} - \boldsymbol{\Sigma}_{um,k}\boldsymbol{\Sigma}_{mm,k}^{-1}\boldsymbol{\Sigma}_{um,k}^{T}, \tag*{(2)}

and

ωk∣xm=ωk⋅ϕ(xm∣μm,k,Σmm,k)∑lωl⋅ϕ(xm∣μm,l,Σmm,l)(3)\omega_{k \mid x_m} = \frac{\omega_k \cdot\phi(x_m \mid\boldsymbol{\mu}_{m,k},\boldsymbol{\Sigma}_{mm,k})}{\sum_l \omega_l \cdot\phi(x_m \mid\boldsymbol{\mu}_{m,l},\boldsymbol{\Sigma}_{mm,l})} \tag*{(3)}

[50, 51, 52] In these equations, μ\boldsymbol{\mu}, Σ\boldsymbol{\Sigma}, ω\omega, and x\boldsymbol{x} represent the mean, covariance, weights, and the known value, respectively. The uu and mm subscripts stand for the unknown and measured data while kk and ll stand of the kthk^{\mathrm{th}} and lthl^{\mathrm{th}} Gaussian component. Note that, for the Single Gaussian Model (SGM) there is only one Gaussian component and the weight will therefore be 1. Lastly, ϕ\phi is a Gaussian Probability Density Function (PDF) with as mean and covariance the parameters after the bar (|) and this is measured at the location of the parameter before the bar.

These equations can be rewritten for the 3D-retinal curvature SGM where the axial length is measured and the radii are the unknown variables. In this case, the measured value is the axial length (AL) with the mean (μAL\mu_{AL}) and the variance (σAL2\sigma_{AL}^{2}).

The unknown variables are the radii [RNTRISRPA]\begin{bmatrix} R_{NT} \\ R_{IS} \\ R_{PA} \end{bmatrix}, with the means μRi\mu_{R_i} (with i=NT, IS, PA depending on the radii of interest). The covariances are σRi,AL\sigma_{R_i,AL} and σRi,Rj\sigma_{R_i,R_j} between the ithi^{\mathrm{th}} radii and the axial length, and the ithi^{\mathrm{th}} radii and the jthj^{\mathrm{th}} radii, respectively.

Resulting in the following conditional distribution parameters:

\begin{bmatrix} \mu_{R_{NT} \mid AL} \\ \mu_{R_{IS} \mid AL} \\ \mu_{R_{PA} \mid AL} \end{bmatrix} = \begin{bmatrix} \mu_{R_{NT}} \\ \mu_{R_{IS}} \\ \mu_{R_{PA}} \end{bmatrix} + \frac{(AL-\mu_{AL})}{\sigma_{AL}^{2}} \begin{bmatrix} \sigma_{R_{NT},AL} \\ \sigma_{R_{IS},AL} \\ \sigma_{R_{PA},AL} \end{bmatrix}, \tag*{(4)}\end{bmatrix}
ΣR∣AL=[σRNT∣AL2σRNT,RIS∣ALσRNT,RPA∣ALσRNT,RIS∣ALσRIS∣AL2σRIS,RPA∣ALσRNT,RPA∣ALσRIS,RPA∣ALσRPA∣AL2],(5)\Sigma_{R \mid AL} = \begin{bmatrix} \sigma_{R_{NT} \mid AL}^{2} & \sigma_{R_{NT},R_{IS} \mid AL} & \sigma_{R_{NT},R_{PA} \mid AL} \\ \sigma_{R_{NT},R_{IS} \mid AL} & \sigma_{R_{IS} \mid AL}^{2} & \sigma_{R_{IS},R_{PA} \mid AL} \\ \sigma_{R_{NT},R_{PA} \mid AL} & \sigma_{R_{IS},R_{PA} \mid AL} & \sigma_{R_{PA} \mid AL}^{2} \end{bmatrix}, \tag*{(5)}

where

σRi∣AL2=σRi2−σRi,AL2σAL2,(6)\sigma_{R_i \mid AL}^{2} = \sigma_{R_i}^{2} - \frac{\sigma_{R_i,AL}^{2}}{\sigma_{AL}^{2}}, \tag*{(6)}
σRi,Rj∣AL=σRi,Rj−σRi,ALσRj,ALσAL2.(7)\sigma_{R_i,R_j \mid AL} = \sigma_{R_i,R_j} - \frac{\sigma_{R_i,AL}\sigma_{R_j,AL}}{\sigma_{AL}^{2}}. \tag*{(7)}

Supplement 2

The final mean matrix μ\mu and covariance matrix Σ\Sigma, which is calculated by taking the average of the five models created for each split, are provided in Table S2.1 and Table S2.2, respectively. These values are needed to generate a cohort of realistic retinal curvatures. If this is directly implemented, a non-conditional SGM is implemented. On the other hand, if these values are implemented in combination with Supplement 1 the conditional SGM can be designed (also see demo.mreye.nl/synteyes).

SGMALRNTR_{NT}RISR_{IS}RPAR_{PA}
μ\mu [mm]23.8511.8911.6610.55

Table S2.1. Mean matrix (μ\mu) for the Single Gaussian Model. The AL, RNTR_{NT}, RISR_{IS}, and RPAR_{PA} represent the axial length and the retinal radii in the nasal-temporal, inferior-superior, and posterior-anterior direction, respectively.

SGM Σ\SigmaALRNTR_{NT}RISR_{IS}RPAR_{PA}
AL1.820.560.420.81
RNTR_{NT}0.560.360.190.18
RISR_{IS}0.420.190.240.16
RPAR_{PA}0.810.180.160.48

Table S2.2. Covariance matrix (Σ\Sigma) for the Single Gaussian Model. The AL, RNTR_{NT}, RISR_{IS}, and RPAR_{PA} represent the axial length and the retinal radii in the nasal-temporal, inferior-superior, and posterior-anterior direction, respectively.

Supplement 3

The axial length and radii of the retinal curvature are compared over the different splits to investigate if these parameters are similar over the splits. After that, the resulting Single Gaussian Models and Multi Gaussian Mixture Models for the individual splits are provided.

Training datasetSplit 1Split 2Split 3Split 4Split 5
Split size7575757576
AL [mm]
Median [mm]23.5623.5623.5723.5623.62
Inter Quartile Range [mm]1.621.341.581.431.89
RNTR_{NT} [mm]
Median [mm]11.8811.8411.8811.8711.89
Inter Quartile Range [mm]0.630.610.610.620.72
RISR_{IS} [mm]
Median [mm]11.6311.6011.6311.6511.65
Inter Quartile Range [mm]0.640.630.670.640.58
RPAR_{PA} [mm]
Median [mm]10.3810.3810.3610.3810.38
Inter Quartile Range [mm]0.930.740.870.791.14

Table S3.1. For each split in the training dataset, the median and interquartile range are provided. The AL, RNTR_{NT}, RISR_{IS}, and RPAR_{PA} represent the axial length and the retinal radii in the nasal-temporal, inferior-superior, and posterior-anterior direction, respectively. It can be seen that these values are very similar over the different splits. This was also verified with the Mann-Whitney U test which did not show a significant difference between the splits.

Validation datasetSplit 1Split 2Split 3Split 4Split 5
Split size1919191918
AL [mm]
Median [mm]23.5923.7023.5923.6623.31
Inter Quartile Range [mm]1.002.061.781.791.14
RNTR_{NT} [mm]
Median [mm]11.8012.0011.8411.8811.84
Inter Quartile Range [mm]0.640.640.650.640.43
RISR_{IS} [mm]
Median [mm]11.6411.7511.6511.6011.43
Inter Quartile Range [mm]0.520.710.410.650.68
RPAR_{PA} [mm]
Median [mm]10.3610.5610.3810.3510.38
Inter Quartile Range [mm]0.481.141.161.180.56

Table S3.2. For each split in the validation dataset, the median and interquartile range are provided. The AL, RNTR_{NT}, RISR_{IS}, and RPAR_{PA} represent the axial length and the retinal radii in the nasal-temporal, inferior-superior, and posterior-anterior direction, respectively. It can be seen that these values are still similar over the different splits. Although, that the differences are larger due to the small split size, still the Mann-Whitney U test did not show a significant difference between the splits.

Violin plots of the distribution of the non-conditional SGM for each split

Figure S3.1. Violin plots of the distribution of the non-conditional SGM for each split. The SGM stays stable over the different splits with only minor deviations in median and interquartile range over the different splits.

Violin plots of the components of the non-conditional MGMM for each split

Figure S3.2. Violin plots of the components of the non-conditional MGMM for each split. For each split, all the pink violins represent the same component and the brown violins the other component. The pink components stays stable over the different splits while the brown component depends on the chosen split. Therefore, the brown component might be modelling noise instead of an underlying component and might be overfitting the data.

Supplement 4

The performance for each model separately over each split for the investigated parameters is provided. Each table consists of a different combination of the models with either the training or the validation dataset.

Non-conditional SGM trainingSplit 1Split 2Split 3Split 4Split 5
ALAL [mm]
Δmedian\Delta_{\mathrm{median}} [mm]-0.27-0.21-0.23-0.30-0.39
F-test (p-value)1.03 (0.80)1.06 (0.67)1.01 (0.91)1.07 (0.66)1.05 (0.74)
Overlap0.890.880.910.890.89
RNTR_{\mathrm{NT}} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]0.03-0.02-0.01-0.05-0.05
F-test (p-value)0.99 (0.98)1.01 (0.91)0.98 (0.96)1.01 (0.91)1.00 (0.96)
Overlap0.900.920.900.900.90
RISR_{\mathrm{IS}} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]-0.02-0.03-0.05-0.03-0.04
F-test (p-value)1.02 (0.88)1.05 (0.74)1.00 (0.95)1.00 (0.94)1.02 (0.88)
Overlap0.940.940.940.940.94
RPAR_{\mathrm{PA}} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]-0.19-0.15-0.15-0.16-0.22
F-test (p-value)1.01 (0.92)1.02 (0.85)1.00 (0.95)1.00 (0.94)1.01 (0.93)
Overlap0.860.850.880.870.85

Table S4.1. For each split in training data, the performance of the non-conditional Single Gaussian Model (SGM) was evaluated with the median difference (Δmedian\Delta_{\mathrm{median}}), f-test, and overlap. The ALAL, RNTR_{\mathrm{NT}}, RISR_{\mathrm{IS}}, and RPAR_{\mathrm{PA}} represent the axial length and the retinal radii in the nasal-temporal, inferior-superior, and posterior-anterior direction, respectively. Over the different splits, the metrices show some variation. However, the model still captured the RNTR_{\mathrm{NT}} and RISR_{\mathrm{IS}} distribution accurately, while it had more difficulties to capture the ALAL and RPAR_{\mathrm{PA}}.

Table S4.2 For each split in training data, the performance of the non-conditional Multi Gaussian Mixture Model (MGMM) was evaluated with the median difference (Δmedian\Delta_{\mathrm{median}}), f-test, and overlap. The AL, RNTR_{\mathrm{NT}}, RISR_{\mathrm{IS}}, and RPAR_{\mathrm{PA}} represent the axial length and the retinal radii in the nasal-temporal, inferior-superior, and posterior-anterior direction, respectively. Over the different splits, the metrices show some variation. However, the model captured all parameters accurately with overlap values exceeding 0.9.

Non-conditional MGMM trainingSplit 1Split 2Split 3Split 4Split 5
ALAL [mm]
Δmedian\Delta_{\mathrm{median}} [mm]-0.060.01>-0.01-0.11-0.10
F-test (p-value)1.04 (0.77)1.08 (0.59)1.02 (0.87)1.07 (0.66)1.01 (0.90)
Overlap0.940.920.930.960.94
RNTR_{\mathrm{NT}} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]0.070.070.08-0.020.04
F-test (p-value)1.04 (0.77)1.02 (0.87)1.00 (0.94)1.03 (0.82)0.97 (0.89)
Overlap0.940.900.910.940.94
RISR_{\mathrm{IS}} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]-0.02-0.03-0.05-0.03-0.01
F-test (p-value)0.99 (0.97)1.03 (0.82)0.99 (0.99)1.02 (0.85)1.00 (0.94)
Overlap0.940.940.940.930.97
RPAR_{\mathrm{PA}} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]-0.10-0.01-0.06-0.03-0.05
F-test (p-value)1.01 (0.92)1.07 (0.62)1.02 (0.86)1.00 (0.96)0.99 (0.99)
Overlap0.920.940.940.950.94

Table S4.2.

Table S4.3 For each split in training data, the performance of the conditional Single Gaussian Model (SGM) was evaluated with the median difference (Δmedian\Delta_{\mathrm{median}}), f-test, and overlap. The RNTR_{\mathrm{NT}}, RISR_{\mathrm{IS}}, and RPAR_{\mathrm{PA}} represent the retinal radii in the nasal-temporal, inferior-superior, and posterior-anterior direction, respectively. Over the different splits, the metrices show some variation. However, the model captured the distributions of each split better than the non-conditional counterpart and all overlap values now exceed 0.9.

Conditional SGM trainingSplit 1Split 2Split 3Split 4Split 5
RNTR_{\mathrm{NT}} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]0.040.01-0.01-0.010.01
F-test (p-value)1.02 (0.88)0.99 (1.00)1.02 (0.86)1.02 (0.86)1.02 (0.86)
Overlap0.920.950.910.930.93
RISR_{\mathrm{IS}} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]-0.03-0.01-0.03-0.01-0.02
F-test (p-value)1.02 (0.86)1.01 (0.89)1.02 (0.88)1.00 (0.95)1.02 (0.85)
Overlap0.950.950.950.950.97
RPAR_{\mathrm{PA}} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]-0.13-0.06-0.11-0.06-0.13
F-test (p-value)1.02 (0.88)1.01 (0.89)1.02 (0.88)1.03 (0.80)1.03 (0.80)
Overlap0.920.950.920.930.91

Table S4.3.

Conditional MGMM trainingSplit 1Split 2Split 3Split 4Split 5
RNTR_{NT} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]0.070.060.07-0.010.05
F-test (p-value)1.00 (0.94)0.98 (0.97)1.00 (0.94)1.02 (0.88)1.02 (0.88)
Overlap0.940.910.900.950.94
RISR_{IS} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]-0.04-0.03-0.05-0.02<0.01
F-test (p-value)1.01 (0.89)1.04 (0.75)0.99 (0.98)1.06 (0.67)1.02 (0.86)
Overlap0.940.950.930.930.97
RPAR_{PA} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]-0.09-0.02-0.05-0.01-0.04
F-test (p-value)1.03 (0.84)1.01 (0.89)1.03 (0.83)0.98 (0.96)0.99 (0.99)
Overlap0.920.950.940.920.94

Table S4.4. For each split in training data, the performance of the conditional Multi Gaussian Mixture Model (MGMM) was evaluated with the median difference (Δmedian\Delta_{\mathrm{median}}), f-test, and overlap. The RNTR_{NT}, RISR_{IS}, and RPAR_{PA} represent the retinal radii in the nasal-temporal, inferior-superior, and posterior-anterior direction, respectively. Over the different splits, the metrices show some variation. However, the model captured all parameters accurately with overlap values exceeding 0.9. The conditional MGMM has similar performance as the non-conditional counterpart.

Conditional SGM validationSplit 1Split 2Split 3Split 4Split 5
RNTR_{NT} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]<0.010.04-0.09-0.070.15
F-test (p-value)0.52 (0.10)1.79 (0.04)0.71 (0.40)1.45 (0.20)0.87 (0.78)
Overlap0.840.910.910.810.86
RISR_{IS} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]0.060.03-0.05-0.10-0.11
F-test (p-value)0.76 (0.50)1.19 (0.53)0.35 (0.01)1.12 (0.64)2.64 (<0.01)
Overlap0.940.880.800.900.73
RPAR_{PA} [mm]
Δmedian\Delta_{\mathrm{median}} [mm]-0.13-0.09-0.12-0.130.11
F-test (p-value)0.92 (0.89)0.86 (0.75)1.67 (0.08)0.95 (0.97)0.88 (0.81)
Overlap0.880.890.820.930.93

Table S4.5. For each split in validation data, the performance of the conditional Single Gaussian Model (SGM) was evaluated with the median difference (Δmedian\Delta_{\mathrm{median}}), f-test, and overlap. The RNTR_{NT}, RISR_{IS}, and RPAR_{PA} represent the retinal radii in the nasal-temporal, inferior-superior, and posterior-anterior direction, respectively. Over the different splits, the metrices show large variation. These variations are caused by the small validation datasets. Therefore, these metrices can not be used for these small datasets.

References

  1. [1]Gullstrand A. Zusatz {II}: Brechung der Strahlen im Auge, Abbildungsgesetze erster Ordnung. In: von Helmholtz H, Gullstrand A, von Kries J, Nagel W. Handbuch der physiologischen Optik (3 ed., Vol. 1). Hamburg und Leipzig: Leopold Voss; 1909:299–306.
  2. [2]Le Grand Y. Optique physiologique. Tome 1: La dioptrique de l’œil et sa correction. Paris: Éditions de la Revue d'Optique; 1952:52.
  3. [3]Emsley HH. Visual Optics (Vol. 1). London: Butterworth; 1962:40–42.
  4. [4]Navarro R, Santamaría J, Bescós J. Accommodation-dependent model of the human eye with aspherics. J Opt Soc Am A. 1985;2(8):1273–1281.pubmed.ncbi.nlm.nih.gov/4032096
  5. [5]Rayamajhi A, Helmer M, Hetz M, Neumann, Cornelius. The human eye: From Gullstrand's eye model to ray tracing today. In: Lux junior 2021 15 Internationales Forum für den lichttechnischen Nachwuchs. Ilmenau Media Services; 2021:201–209.DOI
  6. [6]Van Vught L, Luyten GPM, Beenakker JWM. Distinct differences in anterior chamber configuration and peripheral aberrations in negative dysphotopsia. Journal of Cataract and Refractive Surgery. 2020;46:1007–1015.DOI
  7. [7]Henderson BA, Yi DH, Constantine JB, Geneva II. New preventative approach for negative dysphotopsia. Journal of Cataract and Refractive Surgery. 2016;42(10):1449–1455.DOI
  8. [8]Holladay JT, Simpson MJ. Negative dysphotopsia: Causes and rationale for prevention and treatment. Journal of Cataract and Refractive Surgery. 2017;43(2):263–275.DOI
  9. [9]Van Vught L, Dekker CE, Stoel BC, Luyten GPM, Beenakker JWM. Evaluation of intraocular lens position and retinal shape in negative dysphotopsia using high-resolution magnetic resonance imaging. Journal of Cataract and Refractive Surgery. 2021;47:1032–1038.DOI
  10. [10]Van Vught L, Que I, Luyten GPM, Beenakker JWM. Effect of anatomical differences and intraocular lens design on negative dysphotopsia. Journal of Cataract and Refractive Surgery. 2022;48(12):1446–1452.DOI
  11. [11]Van Vught L, Luyten GPM, Beenakker JWM. Peripheral visual field shifts after intraocular lens implantation. Journal of Cataract and Refractive Surgery. 2023;49(12):1270–1274.DOI
  12. [12]Stenstrom S. Investigation of the variation and the correlation of the optical elements of human eyes. Am J Optom Arch Am Acad Optom. 1948;25(10):496–504.
  13. [13]Rozema JJ, Tassignon M-J. The Bigaussian Nature of Ocular Biometry. Optom Vis Sci. 2014;91(7):713–722.DOI
  14. [14]Michael R, Wirkner K, Engel C, Loeffler M, Kirsten T, Rauscher F. Ocular biometry in adults from a population-based study in Germany. Acta Ophthalmologica. 2022;100(S275):Abstract 0262.
  15. [15]Chen X, Huang Y, Chen H, Liu L. Distribution and Characteristics of Ocular Biometric Parameters among a Chinese Population: A Hospital-Based Study. Ophthalmol Ther. 2023;12(4):2117–2131.DOI
  16. [16]Chu Z, Ren Q, Chen M, et al. The relationship between axial length/corneal radius of curvature ratio and stress–strain index in myopic eyeballs: Using Corvis ST tonometry. Frontiers in Bioengineering and Biotechnology. 2022;10:939129.DOI
  17. [17]Hosny M, Aliό JL, Claramonte P, Attia WH, Pérez-Santonja JJ. Relationship Between Anterior Chamber Depth, Refractive State, Corneal Diameter, and Axial Length. Journal of Refractive Surgery. 2000;16(3):336–340.pubmed.ncbi.nlm.nih.gov/10832983
  18. [18]Chen MJ, Liu YT, Tsai CC, Chen YC, Chou CK, Lee SM. Relationship between central corneal thickness, refractive error, corneal curvature, anterior chamber depth and axial length. Journal of the Chinese Medical Association. 2009;72(3):133–137.
  19. [19]Jivrajka R, Shammas MC, Boenzi T, Swearingen M, Shammas HJ. Variability of axial length, anterior chamber depth, and lens thickness in the cataractous eye. Journal of Cataract and Refractive Surgery. 2008;34(2):289–294.DOI
  20. [20]Gaurisankar ZS, van Rijn GA, Lima JEE, et al. Correlations between ocular biometrics and refractive error: A systematic review and meta-analysis. Acta Ophthalmol. 2019;97(8):735–743.DOI
  21. [21]Rozema JJ, Atchison DA, Tassignon MJ. Statistical eye model for normal eyes. Investigative Ophthalmology and Visual Science. 2011;52(7):4525–4533.DOI
  22. [22]Rozema JJ, Rodriguez P, Navarro R, Tassignon MJ. Synteyes: A higher-order statistical eye model for healthy eyes. Investigative Ophthalmology and Visual Science. 2016;57(2):683–691.DOI
  23. [23]Farzanfar A, Lockett-Ruiz V, Navarro R, Koppen C, Rozema JJ. The influence of variations in ocular biometric and optical parameters on differences in refractive error. Ophthalmic and Physiological Optics. 2024;44(5):1000–1009.
  24. [24]Fischer J, van der Merwe J, Vandenheever D. The influence of eye model parameter variations on simulated eye-tracking outcomes. Journal of Eye Movement Research. 2023;16(3):1.DOI
  25. [25]Rozema JJ, Rodriguez P, Ruiz Hidalgo I, Navarro R, Tassignon M-J, Koppen C. SyntEyes KTC: higher order statistical eye model for developing keratoconus. Ophthalmic and Physiological Optics. 2017;37(3):358–365.DOI
  26. [26]Rozema JJ, Hastings GD, Marsack J, Koppen C, Applegate RA. Modeling refractive correction strategies in keratoconus. Journal of Vision. 2021;21(10):18.DOI
  27. [27]Escudero-Sanz I, Navarro R. Off-axis aberrations of a wide-angle schematic eye model. J Opt Soc Am A Opt Image Sci Vis. 1999;16(8):1881–1891.
  28. [28]Shahiri M, Jóźwik A, Asejczyk M. Opto-mechanical self-adjustment model of the human eye. Biomed Opt Express. 2023;14(5):1923–1944.DOI
  29. [29]Van Vught L, Shamonin DP, Luyten GPM, Stoel BC, Beenakker JWM. MRI-based 3D retinal shape determination. BMJ Open Ophthalmology. 2021;6(1):000855.DOI
  30. [30]Kneepkens SCM, Marstal K, Polling JR, et al. Eye Size and Shape in Relation to Refractive Error in Children: A Magnetic Resonance Imaging Study. Invest Ophthalmol Vis Sci. 2023;64(15):41.DOI
  31. [31]Atchison DA, Pritchard N, Schmid KL, Scott DH, Jones CE, Pope JM. Shape of the Retinal Surface in Emmetropia and Myopia. Investigative Ophthalmology & Visual Science. 2005;46(8):2698–2707.pubmed.ncbi.nlm.nih.gov/16043841
  32. [32]Verkicharla PK, Mathur A, Mallen EAH, Pope JM, Atchison DA. Eye shape and retinal shape, and their relation to peripheral refraction. Ophthalmic and Physiological Optics. 2012;32:184–199.DOI
  33. [33]Kneepkens SCM, Van Vught L, Polling JR, Klaver CCW, Tideman JWL, Beenakker JWM. A Novel MRI-Based Approach to Peripheral Refraction and Prediction of Myopia Progression. American Journal of Ophthalmology. 2025;278:239–249.DOI
  34. [34]Haasjes C, Vu THK, Beenakker JWM. Patient-specific mapping of fundus photographs to three-dimensional ocular imaging. Med Phys. 2025;52(4):2330–2339.DOI
  35. [35]Beenakker JWM, van Rijn GA, Luyten GPM, Webb AG. High-resolution MRI of uveal melanoma using a microcoil phased array at 7 T. NMR in Biomedicine. 2013;26(12):1864–1869.DOI
  36. [36]Hawkins DM, Basak SC, Mills D. Assessing model fit by cross-validation. In: Journal of Chemical Information and Computer Sciences. 2003:579–586.pubmed.ncbi.nlm.nih.gov/12653524
  37. [37]Scott DW. Multivariate density estimation: theory, practice, and visualization. John Wiley & Sons; 2015.
  38. [38]Haasjes C, Kramer H, Beenakker JWM. Visisipy: open and accessible vision simulations in Python. Optica Open Preprint. 2026. https://doi.org/10.1364/opticaopen.34056642.
  39. [39]Nguyen HG, Sznitman R, Maeder P, et al. Personalized Anatomic Eye Model From T1-Weighted Volume Interpolated Gradient Echo Magnetic Resonance Imaging of Patients With Uveal Melanoma. International Journal of Radiation Oncology Biology Physics. 2018;102(4):813–820.
  40. [40]Peel D, McLachlan GJ. Robust mixture modelling using the t distribution. Statistics and Computing. 2000;10(4):339–348.DOI
  41. [41]Lin TI. Maximum likelihood estimation for multivariate skew normal mixture models. Journal of Multivariate Analysis. 2009;100(2):257–265.DOI
  42. [42]Pavlides A, Agou VD, Christopoulos DT. Non-parametric kernel-based estimation and simulation of precipitation amount. Journal of Hydrology. 2022;612:127988.DOI
  43. [43]Artal Soriano P, Ginis H, Barrera Verdejo M, Villegas Ruiz E, Prieto Corrales P, Tabernero de Paz J, inventors; Voptica SL, assignee. Intraocular lens and methods for optimization of depth of focus and the image quality in the periphery of the visual field. US patent US20200383775A1. 2020.
  44. [44]Ghaderi H, García MJ, Koppen C, Rozema J. Opto-biomechanical simulation of keratoconus development in emmetropic eyes. J Optom. 2026;19(1):100587.DOI
  45. [45]Boszczyk A, Dębowy F, Jóźwik A, Daghian A, Siedlecki D. Complexity of crystalline lens wobbling investigated by means of combined mechanical and optical simulations. Biomed Opt Express. 2023;14(6):2465–2477.DOI
  46. [46]Wulff J, Heverhagen JT, Zink K, Kawrakow I. Investigation of systematic uncertainties in Monte Carlo-calculated beam quality correction factors. Phys Med Biol. 2010;55(16):4481–4493.DOI
  47. [47]Ntatsis K, Christaras D, Artal P, Ginis H. Deep learning ocular aberration retrieval from simulated retinal images under straylight. Biomed Opt Express. 2025;16(7):2709–2718.DOI
  48. [48]Zhang H, Zhao J, Chen H, Zhang Z, Yin C, Wang S. Large-Dynamic-Range Ocular Aberration Measurement Based on Deep Learning with a Shack–Hartmann Wavefront Sensor. Sensors. 2024;24(9):2728.DOI
  49. [49]Varadarajan AV, Poplin R, Blumer K, et al. Deep Learning for Predicting Refractive Error From Retinal Fundus Images. Investigative Ophthalmology & Visual Science. 2018;59(7):2861–2868.DOI
  50. [50]Gilardi N, Bengio S, Kanevski M. Conditional Gaussian Mixture Models for environmental risk mapping. Proceedings of the 12th IEEE Workshop on Neural Networks for Signal Processing. 2002:777–786.DOI
  51. [51]Przewięźlikowski M, Śmieja M, Struski Ł. Estimating Conditional Density of Missing Values Using Deep Gaussian Mixture Model. In: 27th International Conference on Neural Information Processing. Springer International Publishing; 2020:220–231.DOI
  52. [52]Chehade A, Savargoankar M, Krivtsov V. Conditional Gaussian mixture model for warranty claims forecasting. Reliability Engineering & System Safety. 2022;218:108180.DOI

Paper details

vlnc:W_awr3wq87
Version
v1 · latest
Published
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents