Sex and age estimation with corneal topography parameters by using machine learning algorithms and artificial neural networks

Abstract

Background

The aim of this study, which was based on this hypothesis, was to estimate sex and age by using a machine learning algorithm (ML) and artificial neural networks (ANN) with parameters obtained from the eyeball. The study was conducted on corneal topography images of 155 women and 155 men aged between 6 and 87 who did not have surgical intervention or pathology in their eyeballs. In the study, the individuals were divided into four different age groups 6–17, 18–34, 35–55, and 56–87. Sex and age estimation was carried out by using the numerical data of parameters obtained as a result of corneal topography imaging in ML and ANN inputs.

Results

As a result of our study, in sex determination, a 0.98 accuracy rate (Acc) was obtained with the logistic regression algorithm, one of the ML algorithms, and 0.94 Acc was obtained with the MLCP model, one of the ANN algorithms; in age estimation, 0.84 Acc was obtained with RF algorithm, one of the ML algorithms. With the SHAP analyzer of the Random Forest algorithm, through which the effects of parameters on the overall result are evaluated, the parameter that made the highest contribution to sex estimation was found to be corneal volume, and the parameter that made the highest contribution to age estimation was found to be pupil Q parameter.

Conclusion

As a result of our study, it was found that parameters obtained from the eyeball showed a high accuracy in sex and age estimation.

Background

Sex and age estimation are very important for forensic, surgical, clinical medical sciences, and retail sales and marketing sectors. Knowing the age and sex of the individual is important in postmortem identification in forensic sciences and preventing child pornography; in the determination of age in forensic cases; in the determination of the risk of diseases such as glaucoma, amblyopia, cataract, hyperopia, myopia, etc. and administering the right surgical intervention in surgical and clinical sciences; and in the presentation of products to right masses in sales and marketing sectors (Bakici et al. 2021; Huynh et al. 2020; Machado et al. 2017; Rüfer et al. 2010; Secgin et al. 2022; Senol et al. 2023; Snellingen et al. 2001; Xiao et al. 2015)).

In diagnosis and treatment processes of eye diseases (keratoconus, corneal ectasia, cataract, etc.), corneal topography devices are frequently used. These devices provide detailed information about the eye by mapping the anterior corneal part of the eye extensively. It is reported that the anterior segment of the eye is measured reliably with this device (Kazancı et al. 2016; Uçakhan 2020). Interpretation of this information provided by corneal topography devices requires serious expertise. Although some of these devices have algorithm-based applications to support interpretation, their prevalence is low and there is no universally accepted software (Shanthi et al. 2022).

Using engineering-based algorithms [artificial neural networks (ANN), machine learning algorithms (ML), deep learning] as assistive in the reconstruction and interpretation of medical images, in the detection and segmentation of pathologies, in data analysis, in classifying reports and identifying morphometric data play a critical role in getting accurate results in a short time. Artificial Neural Networks is a mathematical model that can produce results for problems that consist of input, hidden, and output layers. Machine learning algorithms a modern classifiers with 2 subclasses supervised, unsupervised, and reinforced, and can detect the complex relationships between input and output. Although ML and ANN are engineering-based algorithms, they are widely used in many areas of modern medicine due to their accurate and fast decision-making (Currie et al. 2019; Oner et al. 2019; Toy et al. 2022).

The aim of this study is sex and age estimation by analyzing parameters obtained from the eyeball with modern classifiers of ANN and ML.

Methods

Study population and imaging protocol

The study was conducted with the 2023/1310 numbered decision of the Karabuk University local ethics committee. The study was conducted on 155 women and 155 men (310 individuals in total) between the ages of 6 and 87 who had healthy eyeball structures (with no surgical operation or pathology). The patients were grouped into four as 6–17 years of age (group 1), 18–34 years of age (group 2), 35–55 years of age (group 3), and 56–87 years of age (group 4) for analysis of age.

Images of the eyeball were obtained by using a CSO Sirius + (Costruzione Strumenti Oftalmici) Corneal Topography device (Scandicci, Firenze/Italy).

Image processing

Images that met the inclusion criteria and which were taken between 2022 and 2023 by CSO (Sirius +) Corneal Topography device were scanned retrospectively from Hospital Picture Archiving Communication Systems (PACS) and 15 parameters were measured by using the measurement console of the device (Fig. 1).

Fig. 1

CSO Sirius + corneal topography scan image showing the measured parameters

The parameters used in the study and their definitions are given in Table 1.

Table 1 The measured corneal topography parameters and their definitions

Application of machine learning algorithms and artificial neural networks

Python (Version 3.9) programming language and scikit-learn (Version 1.1.1) framework were used in ML and ANN modeling. ML and ANN models were performed by using Monster Abra A7 V12.5 model computer with 8 Gb Ram and i5 operating system. Twenty percent of the data were used as a test set and 80% were used as the training set. In ML modeling, logistic regression (LR), linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), decision tree (DT), and random forest (RF) algorithms were used for sex estimation, while the RF algorithm was used for age estimation. In the ANN model, a multilayer perceptron classifier (MLCP) was used for the estimation of sex. No extraction or cleaning was applied to the raw data in order for ML and ANN models to be realistic. In MLCP, 16 neurons were used in the input layer and 2 neurons were used in the output layer for sex estimation. In addition, 2 hidden layers were preferred (48 neurons were used in the first hidden layer and 32 neurons were used in the second hidden layer). The data were retrained again 100, 200, and 1000 times so that the topologies could reflect the reality. In both models, Accuracy (Acc), Specificity (Spe), Sensitivity (Sen), and F1 score (F1) values were used for (ML-YSA) performance criteria.

$$\text{Acc}=\frac{\text{TP}}{\text{TP}+\text{FN}+\text{FP}+\text{TN}}$$

$$\text{Sen}=\frac{\text{TP}}{\text{TP}+\text{FN}}$$

$$\text{Spe}=\frac{\text{TN}}{\text{TN}+\text{FP}}$$

$$\text{F}1=2\frac{\text{Precision}\times \text{Recall}}{\text{Precision}+\text{Recall}}$$

Equation 1. (TP; True positive, FP; False positive, TN; True negative, FN; False negative).

LR is an extension of basic regression analysis and can model binary variables in a powerful way. DT is one of the oldest ML algorithm models that models the parameters as if they were a tree structure. In this model, there is a root node and a leaf node, and the branching continues until it reaches the lowest node. DT is easy to interpret as it shows the parameters in a tree structure. RF is an ensemble classifier and works like multiple DTs. RF differs from DT in that it evaluates different parts of the training set and is more sensitive. LDA is a model that categorizes different classes, which is often preferred by anthropologists. QDA is a second-order parametric classifier and is superior to LDA. ANNs are so named because they resemble neurons, which are composed of axons and dendrites. The ANN is in the form of an interconnected node (like neurons are connected to each other and neuronal transmission occurs) and these nodes are called input, output, and hidden layers according to their transformation (Curate et al. 2017; Santos et al. 2014; Toy et al. 2022; Uddin et al. 2019).

Statistical analysis

Whether the data fit normal distribution was tested with Anderson Anderson–Darling test. Median, minimum, and maximum values were included in descriptive statistics. Minitab 17 package program was used for basic statistical analyses.

Results

In the study conducted on 310 individuals between the ages of 6 and 87, it was found with the Anderson–Darling test that the parameters were not normally distributed. Table 2 shows the descriptive statistics of the parameters. ThkMIN, KFMAX, corneal volume and Pupil Q parameters median values were higher in women compared to men. KBMAX, AqD, lens rise, HACD, W-W, iridocorneal angle, HACtilt, and AC volume parameters median values were higher in men than women.

Table 2 Descriptive statistics of the measured corneal topography parameters

Of the ML models with which parameters were evaluated in terms of sex, the highest Acc rate was obtained as 0.98 with the LR algorithm. The accuracy of the other algorithms was between 0.76 and 0.92 (Table 3).

Table 3 Performance criteria obtained in terms of sex as a result of machine learning algorithms (for sex)

The confusion matrix obtained as a result of ML modeling is shown in Fig. 2a–e. In the LR algorithm with the highest accuracy rate, all 32 male individuals in the test set and 29 of the 30 female individuals were correctly predicted.

Fig. 2

Confusion matrix table of machine learning algorithms (for sex). a Logistic regression (LR). b Linear discriminant analysis (LDA). c Quadratic discriminant analysis (QDA). d Decision tree (DT). e Random forest (RF)). *The number 0 represents males and 1 represents females, and 62 individuals in the confusion matrix table consist of individuals in the test set (20%)

The overall effect of all parameters in terms of sex was evaluated by using the SHAP analyzer of the RF algorithm and it was found that the corneal volume parameter provided the highest contribution to accuracy (Fig. 3).

Fig. 3

SHAP analyser of the Random Forest algorithm for sex. *Feature 0: Age, 1: ThkMIN, 2: KFMAX, 3: KBMAX, 4: Gullstrand ratio A/P, 5: Gullstrand ratio, 6: Corneal volume, 7: CCT, 8: AqD, 9: Lens rise, 10: HACD, 11: W-W, 12: Iridocorneal angle, 13: HACtilt, 14: AC Volume

With the MLCP model, in which parameters were evaluated in terms of sex, the highest accuracy rate of 0.94 was obtained as a result of 500 trainings (Table 4). In 100- and 1000-times training, an accuracy of 0.92 was obtained.

Table 4 Performance criteria of artificial neural network model (for sex)

The confusion matrix obtained as a result of ANN modeling is shown in Fig. 4. Twenty-eight of the 32 male and all of the 30 female individuals in the test set in the 500-fold learning were correctly predicted.

Fig. 4

Confusion matrix table of artificial neural network model (for sex). a 100 times training. b 500 times training. c 1000 times training)*The number 0 represents males and 1 represents females, and 62 individuals in the confusion matrix table consist of individuals in the test set (20%)

In age estimation with ML models, an Acc rate higher than 0.75 was obtained only with the RF algorithm, and the rate of other algorithms was ignored because they were lower than 0.75. With the RF algorithm, 0.84 Acc, 0.84 Spe, 0.84 Sen, and 0.83 F1 performance rates were obtained. The confusion matrix obtained as a result of the RF algorithm is shown in Fig. 5.

Fig. 5

Confusion matrix table of machine learning algorithm model (for age). *Number 0 represents the 1st age group (6–17 years), number 1 represents the 2nd age group (18–34 years), number 2 represents the 3rd age group (35–55 years), number 3 represents the 4th age group (56–87 years) and 62 individuals in the confusion matrix table consist of individuals in the test set (20%)

The overall effect of all parameters in terms of age was evaluated by using the SHAP analyzer of the RF algorithm and it was found that the Pupil Q parameter provided the highest contribution to accuracy (Fig. 6).

Fig. 6

SHAP analyzer of Random Forest algorithm for age * Feature 6: Pupil Q, 14: AC Volume, 10: HACD, 9: Lens rise, 5: Corneal volume, 8: AqD, 0: ThkMIN, 2: K^B_MAX, 12: Iridocorneal angle, 3: Gullstrand ratio A/P, 1: K^F_MAX, 11: W-W, 4: Gullstrand ratio P/A, 13: HACtilt, 7: CCT

Discussion

In this study which estimated sex and age with parameters obtained from the eyeball by using ML and YSA methods, an accuracy rate between 0.76 and 0.98 was obtained in sex estimation with ML algorithms, an accuracy rate of 0.84 was obtained in age estimation with RF algorithm and an accuracy rate between 0.92 and 0.94 was obtained in sex estimation with MLCP model. With the SHAP analyzer, it was found that the parameter that made the highest contribution to sex estimation was Corneal volume, while the parameter that made the highest contribution to age estimation was pupil Q.

The eye is the center of the visual sense, about 1 inch in size, which opens the individual to the outside world. The eye receives and focuses incoming rays, converts light energy into chemical and electrical energy, and transmits this energy to the brain (Özer et al. 2016).

The structures that make up the face (eyes, mouth, nose, eyebrows) show differences depending on age and sex. These differences are among the important topics of anthropology. Among the structures that make up the face, the eye contributes greatly to the general aesthetics of the face and its sex and age-related changes (Yasaswini Paladugu 2023).

Due to normal aging of the eye, the corpus vitreum in the eye may lose its jelly-like consistency and become more liquid, the muscles may lose their tense state, and cell death may occur in the cornea and retina. As a result of these conditions, vision loss, focusing problems, decreased sharpness in the eye, narrowing of the visual field, and dry eye may occur. Environmental factors, food habits, and genetic structure affect this normal aging of the eye (Chader et al. 2013; Erdinest et al. 2021).

In eye clinics, devices that come into direct contact with the eye and devices that do not come into direct contact with the eye are used. Devices that do not come into direct contact with the eye are preferred more often since there is no probability of infection and defect. A corneal topography device is a device that provides the opportunity to examine the cornea and anterior chamber in detail with 25 radial sections without direct contact with the eye. Due to this feature, this device is frequently preferred in the examination of individuals with a healthy eye structure and the diagnosis and treatment of individuals with abnormal eye structure in eye clinics (Bayramoğlu et al. 2022; Karakurt 2019; Kazancı et al. 2016; Tekin et al. 2020; Uçakhan 2020; Yıldırım et al. 2013).

In a study they conducted on 107 healthy adults by using a corneal topography device, Kazancı et al. 2016 found corneal volume as 58 ± 3.3 mm³, CCT as 541.5 ± 31.7 μm and reported that corneal volume and CCT decreased with age. In a study they conducted on 30 healthy individuals between the ages of 10 and 53, Tekin et al. 2020 found CCT as 538.9 ± 45.2 µm. In a study they conducted on 50 individuals, Wells et al. 2013 examined CCT with optical pachymetry and found as 523.7 ± 42.3 μm. In a study they conducted on 66 glaucoma, 82 ocular hypertension patients, and 40 controls, they found the mean CCT value as 553.4 ± 37 µm and reported that there were no statistically significant differences between sexes. In a study they conducted on 28 healthy, 28 atopic dermatitis patients, Yıldırım et al. 2013 found CCT in the healthy group as 565.1 ± 22 µm. In a study they conducted on 52 healthy individuals, Gonzalez-Perez et al. 2018 found the mean CCT as 550.4 ± 30.5 µm. In our study, the median value of the corneal volume parameter was 56.30 mm³ in men and 57.90 mm³ in women; the median value of the CCT parameter was 0.55 in men and women. In addition, it was found that the highest contribution among the parameters we used in sex estimation was by corneal volume and this parameter was the fifth in age estimation. These studies in the literature support our results morphometrically.

In a study they conducted on 303 healthy individuals between the ages of 10 and 70, Karakurt et al. 2019 grouped individuals into 7 according to their ages; they found AC volume parameter was between 146.07 ± 48.30 mm³ and 215.27 ± 62.68 mm³ according to age groups, AC volume parameter was reversely correlated with age and it did not have a significant correlation with sex. In a study they conducted on 73 pediatric individuals, Bayramoğlu et al. 2022 found the AC volume value as 159 ± 25 mm³. In our study, we found median AC volume as 178.89 mm³ in men and as 164.63 mm³ in women and we found that it had the 10th effect among all parameters we used to determine sex and the 2nd effect in determination of age. This shows that the AC volume parameter is an important variable in terms of age.

In a study they conducted on 112 healthy individuals between the ages of 15 and 65, Kalaycı et al. 2014 found iridocorneal angle as 41.55 ± 6.39°, CCT as 532.11 ± 36.76 μm, AC volume as 160.43 ± 43.09 mm³. In our study, we found the median iridocorneal angle value as 49.00° in men and 47.30° in women, the median CCT value as 0.55 in men and women, the median AC volume as 178.89 mm³ in men and 164.63 mm³ in women. The literature supports our study morphometrically. In our study, the iridocorneal angle parameter contributed 10th in determining sex and 9th in determining age.

Conclusions

In this study, where sex and age were estimated with parameters obtained from the eyeball, very high accuracy rates were obtained with ANN and ML models. In this respect, we believe that the results will make great contributions to forensic, basic medical sciences, and sales sectors (eyewear, medicine, aesthetics, etc. based on the eyeball).