3.1. Patient sheet window

The digitization of the patient records is surely conducive to an easily search access. It instantly helps to process the patients’ data, complete and read the documentation of the patients’ status, with a better sharing of information intra-medically.

This window is used as a digitization means of the patient files. Moreover, it creates an easily accessible medical database. The ophthalmologists are able to design a new patient sheet. The patients are presented by their first names, last names, their ages, gender, regions and their telephone numbers. The second part of the patient sheet is kept for the surveying factors of keratoconus, notably the presence of allergies, family history, Down syndrome etc.

In addition, the ophthalmologists are able to look for former patients through their file numbers. If necessary, the ophthalmologists can possibly access the historic of the patient file with "HISTORIC" button in order to get historical records of their last visits, interpret the series of the medical information and detect the disease evolution.

Furthermore, the window contains four buttons, as shown in Figure 2. The first buttons identified as "RESET". The second is for the return to the home page. The "SAVE" button is used to save data in a table designed in excel. Next, the "CLOSE" button, as its name suggest, closes the sheet patient window. Finally, the toolbar offers the possibilities of printing, opening the intended, and saving the considered options.

Figure 2 The Patient sheet window. 

With regard to the above mentioned demonstration, the currently used application is not only a simple means of managing the patient files but also of analyzing the images, which the ophthalmologists feel their importance.

3.2. OCT window

The GUI offers an option to open image files, display input image and manipulated image. Ophthalmologists can view and manipulate images. There is a zoom option, used to focus on the most interesting parts of the images as shown in Figure 3.

Figure 3 The OCT window. 

OCT images are usually contaminated with speckle noise (^{Rabbani et al., 1999}). Different image processing can reduce the effect of this noise and improve image quality (^{Chaari et al., 2020}). This interface will give a very simple way to choose exactly the desired operations for an image:

Besides, the user can also manipulate the contrast and the luminosity in order to improve the illumination uniformity of the image.

Pachymetry is possibly estimated by calculating the distance between the two corresponding points of the lower and the upper boundary of the cornea. The pachymetry value is calculated by the following equation.

d: The distance calculated from processed image in pixels.

R: The resolution of the OCT device (=20µm) (^{Korobelnik, 2019}).

"Reset image" button is used to restore the original image and the "Save image" button next to the "Processed Image" window, which helps the user to reuse the manipulated image.

The "Registration" button automatically displays registration window.

3.3. Registration window

The interface offers an option for comparing two images by registration technique. The registration (^{Glocker et al., 2011}) aims at identifying the geometric transformation between two medical images to:

The applied approach is the iconic registration method with a rigid transformation. This approach is used to align cornea images at various states and acquired at different times, facilitating the therapeutic follow-up.

The iconic methods are based on image content using image intensities directly without any reduction. The mean square function is applied as similarity criterion. Afterwards, this criterion will be optimized by an iterative gradient descent method in order to achieve the optimal transformation

The shown registration window in Figure 4 gives an option to open and display the two images for comparison. With the "REGISTRATION" button, the ophthalmologists can view the results of the registration.

Figure 4 Registration window 

3.4. ORBSCAN window

The user can open an image file and display the ORBSCAN image, which facilitates the parameters input.

These parameters are used to calculate the most used criteria by the ophthalmologists, which are embodied in KISA, Efcarpides, CP-TP, and others, thus allowing a direct classification. The cut-off values, which permit the discrimination between the normal, the suspect keratoconus, and the keratoconus corneas, are used. In fact, a color code helps to easily distinguish between the three classes: the green for the normal cornea, the magenta for the suspected keratoconus, and the red for the keratoconus. The "EVALUATE" button displays the color code on the parameter fields as shown in Figure 5.

Figure 5 The ORBSCAN window 

Similarly, these selected and specified corneal parameters, which are under the direction and supervision of the ophthalmologists, are used through an artificial intelligence technique, so as to automate the detection and the classification of the keratoconus. A screening program, based on an Artificial Neural Network (ANN), has been developed. This program is designed to improve the sensitivity of keratoconus detection. Ten parameters are used to build the algorithm, which could best distinguish keratoconus, suspected keratoconus and normal cornea: Corneal topography patterns, colours number in posterior and in anterior float, KISA index, Irregularity in 3mm area, Irregularity in 5mm area, TP, CP, BSF and EFCARPID criterion.

All the existing features are normalized in order to avoid significant differences between variable magnitudes.

The used normalization is z-scores (^{Tounsi et al., 2022}). It is calculated as follows :

The dataset used in this work consists of 98 labelled patient folders. Each patient folder contains an ORBSCAN corneal image, an OCT image and a clinical report. With reference to the ophthalmologists' experience, all the patient records are classified into three different classes: normal, suspected keratoconus and keratoconus. The dataset is segmented as follows: 60% are exploited for the model training, 20% for its validation, and the remaining 20 % for its testing.

Four classifications are mainly schematized. The first classification is binary; it is formed up with two classes: the normal cornea and keratoconus. It is done by using restrictively the already mentioned ten parameters of the ORBSCAN corneal image. However, the second classification is basically designed by three classes: keratoconus, suspected keratoconus and normal, depending on the ORBSCAN clinical signs. The third classification is achieved by combining the ORBSCAN corneal image features with the measures of the corneal pachymetry derived from the OCT image and the clinical judgement. In this light, the patient’s data is sub-divided into two classes: the normal and the keratoconus. The fourth classification is organized by subsuming the patient’s data under three classes, using the 16 multimodal and hybrid features.

The used ANN model is characterized by a multilayer feed-forward network. It contains three distinct layers, namely the input, the hidden, and the output. The information in the network is propagated from one layer to the other. In this regard, the first classification is detailed. The network input layer consists of 10 input neurons, which receive the 10 input values from the ORBSCAN corneal image and then transmit them to the hidden layer. The neural network is evaluated solely by a single hidden layer. Different numbers of neurons, existing within the hidden layer from 1 to 20, are iteratively tested to obtain the best possible accuracy of both training and validation samples. It occurs that the network, containing 11 neurons in the hidden layer performs slightly better, compared to the other tested values. Each hidden neuron layer receives information from the previous one. It performs the weighed summation, which is shown in equation 4. Subsequently, it transforms such information according to its activation function. Referring to the current study, the hyperbolic tangent activation function is adopted, which sends this response to the neuron output layer as a next step.

Where xi (i = 1,2,3 ....n ): inputs coming to the neuron j; Wi (i = 1,2,3 ....n ): weights; n : number of input to the node j; b : Bia is for neuron j and f : activation function. Y_j: output value, which will be passed to the next layer from the neuron j.

The output layer consists of one neuron, which represents both the normal and the keratoconus classes. Each class takes binary values of either 0 or 1. Accordingly, the sigmoid activation function is used for the output layer.

The weights are randomly initialized. During the training process, they are adjusted in accordance with the Levenberg-Marquardt (LM) optimization method (^{Sapna et al., 2012}). It makes the predicated outputs either close to the target or to the actual outputs of the network.

The mean square error between the targets and the predicated outputs of the network is calculated by the following expression.

N represents the total number of the samples, Y signifies the sample actual value, and Y stands for the sample estimated value.

The squared error decreases as the training progresses. The network training stops when the performance reaches the global minimum of the error function in the best case. However, the training exceeds the maximum number of epochs if it is identified as 1000, or else if it exceeds the maximum duration set at 3600 s.

At the same time, the squared error is also evaluated on the validation basis. It allows to validate the learning of the network and to determine its quality to generalize on other examples, as well.

Figure 6 below represents the evolution of the mean square error during the training, validation and testing in relation to the number of Epochs.

Figure 6 Neural network performance as a function of epoch variation. 

The smaller value of the MSE indicates that the ANN model is well trained and has already fitted into the data in an appropriate way. In this respect, the best validation performance occurs at seventh Epochs, with a mean square error of 0.093.

The same overall network architecture of the first classification is applied to the three others, except for a number of parameters presented in Table 1 just below.

After training the network, the proposed classification models for the early detection of keratoconus are tested and evaluated by calculating the accuracy (^{Sahli, 2020}), which is mathematically expressed as:

As shown in Table 1, these classification models have appropriate accuracy for detecting the keratoconus of 96% and 92%, by considering two and three keratoconus classes, respectively. They are increased to 99% for two keratoconus classes while 94% for the three other keratoconus classes, particularly when they are combined with the measures of the corneal pachymetry obtained from the OCT image and the clinical judgement. In this respect, complete data about the living environment, the family history and the cornea pachymetry from the OCT image could considerably refine the diagnosis

The findings prove that the use of the parameters, which are derived from different devices and equally the corneal techniques such as ORBSCAN and OCT, are actually in combination with the five risk factors namely: allergy, family history, eye rubbing, visual blur and primary optical aid. They are likely to result in the detection of the subclinical keratoconus with a higher accuracy rate. The increased performance of the classification models, when adding risk factor, reveals that their evaluation is significantly relevant.

The most appropriate clinical judgment, which is associated with topography and OCT criteria, is the key for building an intelligent multimodal ensuring an effective diagnosis of keratoconus.

This study is aimed at helping ophthalmologists, particularly novice ones, to make right decisions in the light of accurate diagnoses.

The performance of the classification models presented in this research is comparable to the keratoconus detection approaches displayed in the literature as (^{Cao et al., 2020}; ^{Issarti et al., 2019}; ^{Lavric et al., 2020}). Moreover, this work introduces more efficient overall performance, compared to the deep learning models used by ^{Zéboulon et al. (2020)} and ^{Chen et al. (2021)}. These models illustrate the effectiveness of machine learning algorithms on small datasets vis-à-vis deep learning algorithms.

The software is connected to Microsoft Excel to export data as the Excel file. This application has created database containing information and medical parameters which can be useful for other artificial intelligence applications, ameliorating diagnostic aids and statistics.

3.5. Statistics Window

The estimated prevalence in the general population worldwide is estimated at 1.38 / 1000 (^{Hashemi et al, 2020}). It varies widely from 2.3% in central India (^{Gokhale, 2013}) to 0.0003% in Russia.

A certain identification of the prevalence of keratoconus in TUNISIA with its risk factors is helpful for not only setting up an effective monitoring and screening system but also assessing the impact of this disease on the professional, social and personal life of its carriers.

The understanding of the developmental causes of keratoconus has still been limited so far. The statistics ensure a better study of keratoconus by identifying the most attacked people as well as the most influential risk factors.

As shown in Figure 7, this interface presents the statistics space. The ophthalmologists chose the factor, which are to be taken for the study of the prevalence of the disease. These results show that keratoconus appears positively associated with urban region, notably pollution.

Figure. 7 The statistics window 

The proposed application is converted to an executable file, which can be run independently. This software "keratocode.exe"is tested and used in ophthalmological investigations, with promising results in the detection of keratoconus are obtained.