Received: 5 February 2024 Revised: 10 June 2024 Accepted: 16 June 2024 The Journal of Engineering

DOI: 10.1049/tje2.12410

ORIGINAL RESEARCH

Computer vision for eye diseases detection using pre-trained deep

learning techniques and raspberry Pi

Ali Al-Naji1,2 Ghaidaa A. Khalid1 Mustafa F. Mahmood1 Javaan Chahl2

1
Electrical Engineering Technical College, Middle Abstract
Technical University, Baghdad, Iraq
Early diagnosis of eye diseases is very important to prevent visual impairment and guide
2
School of Engineering, University of South appropriate treatment methods. This paper presents a unique approach that can detect
Australia, Adelaide, South Australia, Australia
numerous eye diseases automatically. Initially, this approach used the pre-trained ImageNet
models that provides various pre-trained models for training the acquired data. The exist-
Correspondence
Ali Al-Naji, Electrical Engineering Technical ing data sets are composed of 645 data images acquired clinically, represented by two
College, Middle Technical University, Baghdad groups of subjects as healthy and others holding the proposed eye defect like cataracts,
10022, Iraq.
foreign bodies, glaucoma, subconjunctival haemorrhage, and viral conjunctivitis. Followed
Email: [email protected]
by comparisons of the pre-trained model’s coefficients and prediction performance. Later,
the first-class execution model is integrated within the Raspberry Pi staging and the real-
time digital camera detection. The evaluation process used the confusion matrix, model
accuracy, precision factor, recall coefficient, F1 score, and the Matthews Correlation Coef-
ficient (MCC). Resulting in the performance of these pre-trained ImageNet models used
in this study represented by 93% (InceptionResNetV2), 90% (MobileNet), 86% (Resid-
ual Network ResNet50), 85% (InceptionV3), 78% (Visual Geometry Group VGG19), and
72% (Neural Architecture Search Network NASNetMobile). The results show that the
InceptionResNetV2 achieved the highest performance. This proposed approach shows its
efficiency and strength by early detection of the subject’s unhealthy eyes through real-time
monitoring in the field of ophthalmology.

1 INTRODUCTION part of the eye (sclera) caused by blood vessels, in the conjunc-
tiva rupturing [5]. Viral conjunctivitis is an inflammation of the
Eye diseases are considered one of the main factors in non-fatal conjunctiva caused by an infection. It leads to redness, itching,
disabling conditions and pose significant issues for the health and excessive tearing [6]. Traditional methods utilized for the
of the eye globally. The World Health Organization (WHO) detection of these diseases, like slit lamp biomicroscopy, optical
reported that the estimated financial losses resulting from the coherence tomography (OCT) and ultrasound biomicroscopy
eye defect caused by diseases are about USD 411 billion [1]. Eye come with limitations. These limitations are known as diffi-
diseases include cataracts, glaucoma, foreign bodies, subcon- culties in accessing tools, problems associated with costs and
junctival haemorrhage, and viral conjunctivitis. Cataracts occur precision in performance. Another limitation has been the need
when the eye’s lenses become cloudy, over time leading to a for expert input for interpretation. Additional issues include the
decrease in clarity and blurry vision [2]. Ocular foreign bodies consumption of time and the ability to monitor in real-time
refer to objects or substances that enter the eye and can cause [7–11]. In consequence, modern, effective tactics have become
discomfort, pain, redness and potential harm to the cornea [3]. a necessity to greatly improve the identification and classifica-
Glaucoma encompasses eye conditions that damage the nerve. tion process of these diseases, helping to reach a better-managed
This damage often occurs due to increased pressure within the system.
eye and can result in vision loss if left untreated [4]. A subcon- The analysis and classification operation that medical images,
junctival haemorrhage is characterized by patches on the white including ophthalmic images, undergo, have been possible by

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is
properly cited.
© 2024 The Author(s). The Journal of Engineering published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology.

J. Eng. 2024;2024:e12410. wileyonlinelibrary.com/iet-joe 1 of 11

https://doi.org/10.1049/tje2.12410
2 of 11 AL-NAJI ET AL.

the contributions of machine learning techniques especially cal myopia. For improving the performance of the models,
deep learning. Deep learning arose in the biomedical sciences techniques like image enhancement and deep features extrac-
as a technique that supported clinical decisions. It achieved tion where implemented. It was reported that the DarkNet19
this role by developing the sensitivity, specificity and objectivity model reached a higher accuracy with 0.936 compared to the
aspects of the decision-making program, thus making a contri- SqueezeNet model that showed a lower accuracy represented
bution in the upgrade of disease monitoring and detection to a by 0.875.
higher standard [12, 13]. This improvement has been noticeable, Another study by Camera et al. (2022) [19] used ten
particularly in recent years, where noteworthy demonstrations different pre-trained ImageNet models (VGG19, VGG16,
of potential have been expressed by the convolutional neural Xception, ResNet50, MobileNet, DenseNet, MobileNet, Incep-
networks (CNNs) based model ImageNet, a pre-trained deep tionResNetV2, InceptionV3, NASNetMobile) to automate the
learning model. An instance of this significant performance detection of glaucomatous papillae using fundus imaging. The
was in the study by Bhowmik et al. (2019) [14] where Ima- highest accuracy achieved was 0.98 with the VGG19 model,
geNet pre-trained models were used, including VGG16 and while the lowest was 0.87 with the Xception model when eval-
InceptionV3, in the classification of fundus images by utilizing uated on a randomized test set. Using the Madrid and Zaragoza
datasets consisting of 4000 fundus images. The outcomes show- test suite for the dataset, the highest accuracy achieved was 0.92
cased an average prediction accuracy of 0.926 for InceptionV3 with VGG19, while MobileNet had the lowest accuracy of 0.77.
and 0.94 for VGG16. Another study by Diaz-Pinto et al. (2019) Another study by Zhou (2022) [20] used the M-ResNet model
[15] operating with the pre-trained ImageNet models package to develop an intelligent system for diagnosing cataracts, sub-
(five pre-trained models), including the VGG16, VGG19, the conjunctival haemorrhage, keratitis, and pterygium. The results
InceptionV3, ResNet50, and Xception by making use of fun- demonstrated the system’s effectiveness, with the best accuracy
dus images to achieve instant glaucoma detection abilities. The reaching 0.875. Another study conducted by Rajee et al. (2023)
results based on five public databases (1707 images) showed that [21] used the InceptionV3 model for the automated prediction
the best accuracy achieved was 0.96 using the Xception model. of diabetic retinopathy, age-related macular degeneration, glau-
Another research by Smaida and Yaroshchak (2020) [16] used coma, haemorrhages, and epiretinal membrane diseases. The
CNN, VGG16 and InceptionV3 models in the classification of study achieved an accuracy of 0.891 using the InceptionV3
eye diseases of varying classes, including diabetic retinopathy, model on a dataset with 3294 images from the multiple dis-
glaucoma, and myopia, using a dataset of 2781 retinal fundus ease dataset (MUD). In a recent investigation by Patil et al.
images. The end results showed that CNN achieved a predic- (2023) [22], an ensemble model that combines Xception, Incep-
tion accuracy of 0.715 (lowest), while InceptionV3 achieved tionV3, and Densenet201 models in parallel was proposed for
the highest accuracy of 0.877 (best). In another research by the detection of multiple diseases, including cataract, pathologi-
Raza et al. (2021) [17], the classification of eye diseases and cal myopia, glaucoma, and diabetic retinopathy. By combining
sensing cataracts, glaucoma and retinal diseases was attained the datasets from the ODIR and the Joint Shantou Interna-
by Inception-V4 deep learning model and digital fundus imag- tional Eye Centre (JSIEC) and applying transfer learning, the
ing (DFI). The main results displaced the high accuracy of the proposed model achieved an accuracy of 0.88. While the studies
suggested method, which was 0.96 from 601 fundus images. mentioned above highlight the potential of artificial intelligence-
Additional research by Khan et al. (2021) [16] presented the based approaches for the automated detection of eye diseases
VGG-19 pre-trained model for detecting cataract showcasing, and assist ophthalmologists with accurate diagnosis and treat-
which detects cataracts from around 800 colours of the fun- ment decisions; however, these studies mostly used fundus
dus images, including 0.97 accuracies. Later, both Gour and images rather than anterior images of the eye in diagnosing
Khanna (2021) [18] worked with four pre-trained models of eye diseases. Relying on fundus images only, which requires
ImageNet, including the ResNet, InceptionV3, MobileNet, and specialized equipment such as a fundus camera to capture high-
VGG16, that were used to detect various diseases of the eye. resolution images of the posterior segment, makes it impractical
These diseases were represented by diabetic retinopathy and to replace traditional methods completely and presents chal-
glaucoma alongside age-related macular degeneration, myopia, lenges in real-time assessment due to limited access to fundus
hypertension, and cataracts. This research has been based on images from digital cameras or Mobile cameras. While it is tech-
the data set of fundus image pairs from Eye Disease Intelli- nically possible to obtain fundus images from digital cameras or
gent Recognition (ODIR), collected from 5000 patients through mobile cameras using specific accessories or adapters, this usu-
numerous medical centres in China. It was reported that the ally involves an additional cost, and the resulting images may
VGG19 model was validated with a higher accuracy repre- show lower quality and resolution, leading to inaccuracies in
sented by 0.871 compared to the ResNet model which achieved the assessment and diagnosis of eye diseases. One notable study
a lower accuracy represented by 0.849. Recently, there were conducted by Siddique et al. (2022) [23] focused on diagnosing
five models represented by DarkNet19, AlexNet, SqueezeNet, eye diseases, such as cataracts, chalazion, and strabismus, using
Dense-Net201, and ResNet18 were trained in Choudry et al. anterior images of the eye instead of fundus images. They used
(2022) [18] study by working on 3726 images of the fundus six pre-trained models (VGG16, VGG19, MobileNet, Xcep-
in order to help in diagnosis of numerous eye diseases. Dis- tion, InceptionV3, and DenseNet121) to recognize and classify
eases like age-related macular degeneration, glaucoma, cataracts, these diseases. Among the models, MobileNet achieved the
diabetic retinopathy, hypertensive retinopathy, and pathologi- highest accuracy at 0.975. This highlights the efficacy of using
AL-NAJI ET AL. 3 of 11

FIGURE 1 Visual examples of eye disease cases included in the dataset.

anterior eye imaging and deep learning models to accurately ticipants before their images were used for research purposes.
diagnose and classify eye diseases, offering potential advances in The dataset used in this research was securely stored to protect
ophthalmic diagnosis. However, this study did not extensively the privacy and confidentiality of the participants.
explore the algorithm’s applicability to other eye diseases and The study included a dataset of around 645 subjects’ eye
imaging modalities. In addition, it is important to note that the images, where only 445 images of them referred to subjects hav-
above studies did not incorporate hardware for real-time assess- ing eye diseases and actually 200 eye images related to subjects
ment, which could be a potential limitation in implementing having healthy eyes. The dataset was collected from the Balad
the algorithm in real-time clinical settings. Therefore, the work Ruz General Hospital and Ibn Al-Haitham Teaching Eye Hos-
of this study based on the pre-trained ImageNet models pack- pital, involving people with different eyes condition visiting the
age using NASNetMobile, VGG19, ResNet50, InceptionV3, hospitals for regular check-ups starting from the second of Jan-
MobileNet, and InceptionResNetV2 pre-trained models is pro- uary till the seventh of July 2023. The collected eye’s images
posed for the detection of various eye diseases like the cataract, were classified under the supervision of experienced ophthal-
presence of the foreign body, the appeared of glaucoma, the mologists by six categories: the normal or healthy group (200
haemorrhage subconjunctival, and lastly viral conjunctivitis images), the cataract group (100 images), the group of foreign
from the dataset of the anterior images, when these limitations body (91 images), the group of glaucoma (79 images), the sub-
are main considerations. The performance of these models in conjunctival haemorrhage group (84 images), and the group of
detecting these diseases is evaluated, and a comparison is made viral conjunctivitis (91 images). The saved images were format-
to recognize the model with high efficiency to be incorpo- ted in the JPEG file format. Some examples of each eye disease
rated within the platform of Raspberry Pi intended for real-time case used in this study are shown in Figure 1.
appraisals.
This paper is organized as follows: Section 2 provides an
overview of the materials and methods used in the study, incor- 2.2 Experimental setup
porating the research ethics, collection of data, the experimental
environment, the hardware and software configuration recalled The procedure allowed the incorporation of the named models
by the pre-processing, the performance of ImageNet models previously trained with the proposed microcomputer (Rasp-
package and finally the process of evaluation and comparison berry Pi), which allowed the implication of the real-time
of metric’s performance. Section 3 presents the study results and captured images. The experimental setup involved placing the
discussion. Finally, section 4 involves the conclusion. subject within the front view from the camera by 0.5 m, as
shown in Figure 2. This programming language, represented by
Python version 3.9 and the platform of Anaconda version 2.3.2,
2 MATERIALS AND METHODS was used in the implementation of the diagnosis process.

2.1 Research ethics and data collection

2.3 Hardware
The study was conducted according to the standards of eth-
ical conduct in the Declaration of Helsinki, which involve The hardware part of the experiment composed of the follow-
humans as participants. The approved ethics were obtained ing components, (1) the Raspberry Pi 4 platform of Model B,
from the Ethics Committee at the Training and Human Devel- (2) the efficient touch screen, and (3) the high-performance dig-
opment Centre, Iraqi Ministry of Health (protocol number: ital camera as shown in Figure 2. The Raspberry Pi 4 Model
2003016-1981). Informed consent was obtained from all par- B, which is a popular microcomputer single-board type, worked
4 of 11 AL-NAJI ET AL.

2.4 Data pre-processing

The images taken were pre-processed as required, such as resiz-

ing, normalization, and the augmentation process, before being
used to prepare and appraise the pre-trained models. The pro-
cess of data augmentation was implemented through de-scaling
to normalize pixel values by performing a rotation with a 5◦
angle to produce various angles of view, then slightly served to
change the image width and height by 5% to improve the perfor-
mance of the model. For lighting variations, the brightness was
attuned within the range of [0.9, 1.1]. Also, a small shear and
zoom (0.05) and zoom represented by 0.05 and 0.05, respec-
tively, were utilized to introduce both the smooth skews and
varies in the scale of the image correspondingly. The previously
described techniques served to increase the dataset’s size and
FIGURE 2 Example of patient eye disorder case from images dataset. improve the robustness of the pre-trained deep learning model
to be robust in handling distinctions frequently located in the
analysis of the image. To ensure the best and most accurate per-
as the core for the proposed hardware system used to provide formance, the images of the eye dataset were split by 70% of
computational power, the required connectivity for real-time them for training purposes and 30% of them for testing pur-
imaging processes, and finally, the interface of deep learning. In poses, and using epochs were selected within 100 to secure the
addition, this type of Raspberry Pi 4 Model B was chosen for the best performance.
processing capabilities of enhancement, the ability to improve The Haar cascade classifier algorithm [25, 26] was used for
performance and the highest memory capacity [24] compared eye detection, in order for extracting and isolating both eyes
to other Raspberry Pi models. The touch screen has been inte- for the process of analysis. The Haar cascade classifier is a
grated into the proposed system of specification represented by machine learning-based approach employing Haar-like features
WaveShare HDMI Screen, and 5-inch with 800 × 480 pixel res- and the AdaBoost algorithm to detect objects in images, partic-
olution. This camera is used for the subject interface and the ularly effective in identifying well-defined features, such as faces,
monitoring of eye’s condition. Both the Kisonli camera (Model eyes, or other objects with distinct shapes, stored in XML files.
No.: U-227, digital zoom (f = 3.85 mm)) were employed to cap- By comparing the obtained subject eyes to the deep learning
ture images of the eye in a real-time environment and compare pre-trained model, it allows for accurately classified and pre-
them with the pre-trained model in order to predict the case of dicted eye defects or healthy eye states. The block diagram of
the eye. the proposed system is shown in Figure 3.

FIGURE 3 The block diagram of the proposed system.

AL-NAJI ET AL. 5 of 11

the frames of the greyscale and RGB. By analysing, the region

of interest (ROI) was resized to the required input size of the
image aimed at the model’s detection. The region of interest
(ROI) was resized and converted to a matrix format to be used
for prediction.
The prediction model was fed by the ROI based on the pre-
trained detection model. The predicted label, which represents
the state of the subject’s detected eye, was recalled from the cate-
gory labels based on the highest probability. Then, the projected
label was displayed on the frame. The frame was displayed with
drawn rectangles and projected labels, allowing observing the
detection process in real-time. Once the round ends, the captur-
ing process by video device is released, followed by closing of
all the windows.

2.5 ImageNet pre-trained deep learning

models
The pre-trained ImageNet models package was used in this
study to detect eye diseases. These pre-trained models have been
trained to a large extent on the large-scale ImageNet dataset,
which contains millions of annotated images of various object
categories. The following models were used:

∙ NASNetMobile: The NASNetMobile architecture, intro-

duced by Zoph et al. [27], is a developed CNN model that
has gained recognition for its exceptional performance. The
NASNetMobile architecture consists of a total of 53 convolu-
tional layers. These layers are structured by numerous blocks
and repeated structures of cells, creating a deep network for
effective feature extraction and representation learning. The
parameters of trainable were around 5.3 million. The size of
the image that is referred to as an input to the NASNetMo-
bile architecture is 224 × 224 × 3. The specific configuration
and architectural details of the NASNetMobile model can be
found in [27].
∙ VGG19: Firstly, Simonyan and Zisserman introduced this
VGG19 architecture as reported by [28], work as the CNN
model that developed to be recognized for its unique per-
FIGURE 4 The real-time eye diseases detection system flowchart using formance. It consists of 19 layers, including 16 convolutional
the pre-trained deep learning model.
layers and three fully connected layers. The structural design
of VGG19 is identified for firstly its deep structure and sec-
The flowchart of the proposed system is shown in Figure 4, ondly its homogeneous design, therefore the multiple 3 × 3
which illustrates the series of strategies used with Python code. convolutional layers have been stacked to shape their deeper
Starting by loading the pre-trained model and then Haar cascade demonstrations. This model has about 144 million trainable
classifier used for the process of eye detection. The category parameters. Usually, the size of the input image that is used
labels were outlined to stand for the various eye diseases. The for the model of VGG19 is 224 × 224 × 3. The configura-
capturing of video apparatus of video capturing was opened to tion details of the model called VGG19 described in details
get frames of image for the processing. Where the frame was in [28].
taken continuously from the video capture device. Then for the ∙ ResNet50: The architecture of ResNet50 model, introduced
purpose of eye detection, it was converted firstly to an RGB by He et al. [29], which was a CNN developed model
colour space and then to grayscale frame. Later, the Haar classi- that related to the Residual Network family (ResNet) . This
fier used the cascades with the greyscale frame for the detection ResNet50 model has had an outstanding performance in dif-
of the eyes. During the process of eye’s detection, a rectangle ferent tasks of computer vision and it has been considered
was sketched about it on the displayed frame. After that, the as a popular option for the process of both the image clas-
region of interest (ROI), including the eyes, was extorted from sification and the image feature extraction. This deep model
6 of 11 AL-NAJI ET AL.

totally comprises of 117 layers, compromising the convolu- This factor represents by the proportion of the TP predictions
tional layers, pooling layers, fully connected layers, and lastly amongst all the actual positive cases. Furthermore, the F1 score
the shortcut connections layers. The trainable parameters is a harmonic way of accuracy and recovery, by capturing the
were around 25.6 million approximately. Usually, the size of balance between both accuracy and recovery. This factor served
the input image that used for the model of ResNet50 is as a single metric to evaluate the model performance instanta-
224 × 224 × 3. The configuration details of the model called neously in both accuracy and recall. Finally, the factor called by
ResNet50 described in details in [29]. MCC is represented by a quality measure for the binary classifi-
∙ InceptionV3: The architecture of InceptionV3 model, intro- cation models that counts the predictions of the TP, TN, FP, and
duced by Szegedy et al. [30], is a CNN model that developed FN. By evaluating the overall agreement between both the pre-
and modified to work effectively and professionally with a dicted label and true label, which ranging from −1 to +1, since
complex image tasks. This deep model is composed of 316 +1 shows a perfect prediction, while 0 indicates a random pre-
layers, represented by the convolutional layers, the pooling diction, and −1 represents the complete disagreement. All the
layers, and the startup modules, acknowledging the represen- outlined parameters have been expressed as follows [33–35]:
tation of the robust feature. The trainable parameters of this
model was around 23.9 million, and the size of its input image (TP + TN )
where represented by 299 × 299 × 3 The configuration details Accuracy = (1)
(TP + TN + FP + FN )
of the model called InceptionV3 described in details in [30].
∙ MobileNet: The architecture of MobileNet model intro- Precision =
TP
(2)
duced by Sandler et al. [31], is a CNN model that developed (TP + FP )
and modified to work effectively and professionally with TP
a complex image tasks. This deep model is composed of Recall = (3)
(TP + FN )
155 layers, represented by the depth-wise separable convo-
lutional, the bottleneck, and the linear layers. The trainable 2 × (Precision × Recall )
F 1 Score = (4)
parameters of this model were around 3.5 million, forming it (Precision + Recall )
as lightweight and efficient model. The size of its input image (TP × TN − FP × FN )
were represented by 224 × 224 × 3. The configuration details MCC = √
of the model called InceptionV3 described in details in [31]. ((TP + FN ) (TN + FP ) (TP + FP ) (TN + FN ))
∙ InceptionResNetV2: The mode design of the Inception- (5)
ResNetV2 presented by Szegedy et al. [32], characterized
as a highly processed CNN model that joins the model’s
strength of both the Inception and the ResNet architectures. 3 EXPERIMENTAL RESULTS AND
The deep learning model consists of 825 layers, presenting DISCUSSION
a design of complex network that incorporates the convolu-
tional layers, the pooling layers, the inception modules, and This section presents the performance of the previously illus-
the residual connections. This model has trainable param- trated pre-trained package of models (the NASNetMobile
eters around 55.9 million with the size of an input image model, the model of VGG19, InceptionV3, the model of
appearing as 299 × 299 × 3. The configuration details ResNet50, the MobileNet model, and the InceptionRes-NetV2
of the model called InceptionV3 are described in detail model) used for the purpose of identifying the state or situation
in [32]. of the left and right eye from the subject’s eye image dataset
collected from the hospital. The performance of these model
metrics, represented by accuracy, precision, recall, F1 score and
2.6 Evaluation metrics MCC, was calculated and summarized in Table 1 and Figure 5.

Generally, the performance of the previously outlined deep

learning pre-trained models was evaluated by various evaluation TABLE 1 Performance metrics for detecting eye diseases using different
metrics, incorporating the confusion matrix, the accuracy, the pre-trained models.
precision, the recall, the F1 score, and the MCC. Initially, the
Matthews
confusion matrix used to provide a tabular demonstration of F1 correlation
the deep learning model’s performance, presenting the number Model Accuracy Precision Recall score coefficient
of the true positive (TP), the true negative (TN), the false pos-
NASNetMobile 0.72 0.68 0.68 0.67 0.65
itive (FP), and the false negative (FN) predictions. Additionally,
VGG19 0.78 0.75 0.74 0.74 0.72
the accuracy factor which defined as the ratio of the correct pre-
dictions to the total number of performed predictions made. InceptionV3 0.85 0.83 0.83 0.83 0.82
Also, the precision factor is identified as the positive predic- ResNet50 0.86 0.86 0.83 0.84 0.83
tive value, which measures the proportion of TP predictions of MobileNet 0.90 0.91 0.88 0.89 0.88
the model amongst all the positive predictions. In addition, the InceptionResNetV2 0.93 0.94 0.92 0.93 0.92
recall factor used which called as the sensitivity or the TP rate.
AL-NAJI ET AL. 7 of 11

FIGURE 5 Performance metrics for eye disorder detection at different pre-trained models.

From Table 1 and Figure 5, it is clearly illustrated that the for the NASNetMobile and VGG19 models is illustrated in
InceptionResNetV2 model showcased an outstanding perfor- Figure 6.
mance compared to other models by consistently achieving The confusion matrix for the InceptionV3 and ResNet50
the highest scores across all metrics. It achieved an accuracy models is shown in Figure 7.
score of 0.93, precision of 0.94, recall of 0.92 and F1 score of The confusion matrix for the MobileNet and InceptionRes-
0.93. According to these results, the InceptionResNetV2 model NetV2 models is shown in Figure 8.
excels in accurately detecting eye diseases. Following closely Finally, the performance of the best model represented
behind, the MobileNet and ResNet50 models also displayed by InceptionResNetV2 through training and validation shows
quality performances. They have proved their effectiveness in accuracy and loss trends across 100 epochs for eye diseases
eye disease detection based on their exhibited high scores on detection, as shown in Figure 9. From the last mentioned figure,
all metrics. Furthermore, the VGG19 and InceptionV3 mod- it is evident that the accuracy of training steadily increases
els showed moderately high scores, but on the other hand, over the epochs, reaching a maximum accuracy of approxi-
the NASNetMobile model performed inadequately across all mately 96%, compared to the accuracy of validation, which also
metrics. increases, getting a maximum value of about 85%. Furthermore,
In addition, it is observed from Table 1 and Figure 4 that the training and validation loss curves appeared to have a similar
the InceptionResNetV2 model achieved the most excellent nor- trend. The training loss gradually decreases as the model learns,
malized MCC score of 0.92, an additional teller of its brilliant while the validation loss remains consistently lower, indicating
eye disease detection. Its performance was almost matched by good generalization. This deep learning model, characterized
yet another strong performance from the MobileNet model, by InceptionResNetV2 was learning efficiently to classify the
which has an MCC score of 0.88. Other models with relatively training data of eye disease and circulates appropriately well
high MCC scores are ResNet50 and InceptionV3 with 0.83 and for both the new and unseen data throughout the validation
0.82, respectively, stipulating their efficient eye disease sensing. process.
The model with a slightly lower than higher performance with a The experimental results from the other deep learning mod-
0.72 score, is the VGG19. The NASNetMobile model attained els represented by the NASNetMobile, VGG19, InceptionV3,
the lowest MCC score of 0.65, suggesting that the nature of its ResNet50, MobileNet, and InceptionResNetV2 models show
execution is relatively lower in its eye disease observations. a promising performance in eye disease detection. However,
The confusion matrix’s purpose is to visually depict the the proposed system of the outlined previous models has some
true positives, true negatives, false positives, and false negatives limitations. One limitation is the potential impact of subject
distribution for detecting eye diseases. The confusion matrix movement during imaging, image background, and low-contrast
8 of 11 AL-NAJI ET AL.

FIGURE 6 The confusion matrix illustrating the classification results for detecting eye disorder using (a) the NASNetMobile model and (b) the VGG19 model.

FIGURE 7 The confusion matrix illustrating the classification results for detecting eye disorder using (a) the InceptionV3 model and (b) the ResNet50 model.
AL-NAJI ET AL. 9 of 11

FIGURE 8 The confusion matrix illustrating the classification results for detecting eye disorder using (a) the MobileNet model and (b) the InceptionResNetV2
model.

FIGURE 9 Training and validation performance of the InceptionResNetV2.

10 of 11 AL-NAJI ET AL.

images, resulting in the introduction of artifacts and its effect methodology; validation; visualization; writing—review and
on the accuracy of predictions. In addition, the reflections editing. Javaan Chahl: Funding acquisition; methodology;
of eye-lighting that appeared in the captured images might visualization; writing—review and editing.
pose challenges for accurate classification. Also, the limited
cases in the data set of some eye models’ disease categories CONFLICT OF INTEREST STATEMENT
could affect these models’ generalizability. The proposed sys- The authors declare no conflicts of interest.
tem was also restrained by its inability to assess the severity
or degree of eye disease detected accurately. Thus, it cannot DATA AVAILABILITY STATEMENT
specifically inform the operator about the case severity. This The data is available on request from the corresponding author.
limitation will cause difficulty in the clinical staff’s composing
appropriate decision-making due to insufficient deep informa- ORCID
tion. Moreover, using a single-board computer like Raspberry Ali Al-Naji https://orcid.org/0000-0002-8840-9235
Pi 4 could potentially have limitations in terms of scalability, Ghaidaa A. Khalid https://orcid.org/0000-0001-6270-6445
expandability, and connectivity compared to traditional desktop
setups. REFERENCES
Future studies could address all of these limitations by 1. WHO.: Blindness and vision impairment. https://www.who.int/news-
tackling motion artefacts, improving the techniques of image room/fact-sheets/detail/blindness-and-visual-impairment. Accessed 1st
capturing, and increasing the data set to provide a more diverse July 2023.
representation to quantify the severity of eye diseases. Fur- 2. Delbarre, M., Froussart-Maille, F.: Signs, symptoms, and clinical forms of
cataract in adults. J. Fr. Ophtalmol. 43, 653–659 (2020)
thermore, image background and low-contrast images can be
3. Pandey, A.N.: Ocular foreign bodies: A review. J. Clin. Exp. Ophthalmol.
addressed based on an enhancement approach proposed in [36]. 08, 1â (2017)
Also, the deep learning models can be broadened in their clas- 4. Schuster, A.K., Erb, C., Hoffmann, E.M., Dietlein, T., Pfeiffer, N.: The
sification to other eye diseases, comprising multiple diseases diagnosis and treatment of glaucoma. Dtsch. Arztebl. Int. 117, 225 (2020)
within a single patient. The improvement of the classification 5. Doshi, R., Noohani, T.: Subconjunctival hemorrhage. StatPearls Publishing
LL, Treasure Island, FL (2019)
accuracy of diseases needs extensive training and the inclusion
6. Solano, D., Fu, L., Czyz, C.N.: Viral conjunctivitis. StatPearls Publishing
of images from other classes. Normalizing the deep learn- LL, Treasure Island, FL (2017)
ing model by incorporating a diverse data set will improve 7. Prananda, A.R., Frannita, E.L., Hutami, A.H.T., Maarif, M.R., Fitriyani,
its robustness and effectiveness in accurately classifying eye N.L., Syafrudin, M.: Retinal nerve fiber layer analysis using deep learning
diseases. to improve glaucoma detection in eye disease assessment. Appl. Sci. 13, 37
(2022)
8. Cheung, C.Y., Tang, F., Ting, D.S.W., Tan, G.S.W., Wong, T.Y.: Artificial
intelligence in diabetic eye disease screening. Asia-Pacific J. Ophthalmol. 8,
4 CONCLUSION 158–164 (2019)
9. Paradisa, R.H., Bustamam, A., Mangunwardoyo, W., Victor, A.A.,
Yudantha, A.R., Anki, P.: Deep feature vectors concatenation for eye
This study presented an automated system that showcased a
disease detection using fundus image. Electronics 11, 23 (2021)
computer vision approach. This system is used for detecting var- 10. Abbas, Q.: Glaucoma-deep: detection of glaucoma eye disease on retinal
ious eye diseases (cataracts, foreign bodies, glaucoma, subcon- fundus images using deep learning. Int. J. Adv. Comput. Sci. Appl. 8(6),
junctival for haemorrhage and viral conjunctivitis), alongside 080606 (2017)
the detection of healthy eyes. This was achieved by integrating 11. Subramanian, M., Kumar, M.S., Sathishkumar, V., Prabhu, J., Karthick, A.,
Ganesh, S.S., Meem, M.A.: Diagnosis of retinal diseases based on Bayesian
pre-trained deep learning models from ImageNet, including the
optimization deep learning network using optical coherence tomography
NASNetMobile, VGG19, ResNet50, InceptionV3, MobileNet, images. Comput. Intell. Neurosci. 2022, 1–15 (2022)
and InceptionResNetV2) on the platform of Raspberry Pi. 12. Elangovan, P., Vijayalakshmi, D., Nath, M.K.: Covid-19net: An effective
The performance of these stated deep learning algorithms and and robust approach for covid-19 detection using ensemble of convnet-
advanced embedded systems was illustrated in their experimen- 24 and customized pre-trained models. Circuits Syst. Signal Process. 43,
2385–2408 (2023)
tal outcomes. The InceptionResNetV2 and MobileNet achieved
13. Vijayalakshmi, D., Elangovan, P., Nath, M.K.: Methodology for improv-
the highest accuracy, represented by 93% and 90%, respectively, ing deep learning-based classification for CT scan COVID-19 images.
while the NASNetMobile model showed the lowest precision, In: Biomedical Engineering: Applications, Basis and Communications
reaching 72%. This automated diagnostic system will be capable 2450008. World Scientific, Singapore (2024)
of improving the eye healthcare system’s performance. 14. Bhowmik, A., Kumar, S., Bhat, N.: Eye disease prediction from optical
coherence tomography images with transfer learning. In: Proceedings of
the 20th International Conference on Engineering Applications of Neural
AUTHOR CONTRIBUTIONS Networks, EANN 2019, pp. 104–114. Springer, Cham (2019)
Ali Al-Naji: Conceptualization; formal analysis; investiga- 15. Diaz-Pinto, A., Morales, S., Naranjo, V., Köhler, T., Mossi, J.M., Navea,
tion; methodology; project administration; software; vali- A.: CNNs for automatic glaucoma assessment using fundus images: an
extensive validation. Biomed. Eng. Online 18, 1–19 (2019)
dation; visualization; writing—original draft. Ghaidaa A.
16. Smaida, M., Yaroshchak, S.: Bagging of convolutional neural networks for
Khalid: Data curation; formal analysis; investigation; method- diagnostic of eye diseases. Proc. COLINS, 2604, 715–729 (2020)
ology; resources; validation; writing—review and editing. 17. Raza, A., Khan, M.U., Saeed, Z., Samer, S., Mobeen, A., Samer, A.: Classifi-
Mustafa F. Mahmood: Data curation; formal analysis; cation of eye diseases and detection of cataract using digital fundus imaging
AL-NAJI ET AL. 11 of 11

(DFI) and inception-V4 deep learning model. In: Proceedings of the 2021 28. Simonyan, K., Zisserman, A.: Very deep convolutional networks for
International Conference on Frontiers of Information Technology (FIT), large-scale image recognition. In: Proceedings of the 3rd International
pp. 137–142. IEEE, Piscataway, NJ (2021) Conference on Learning Representations (ICLR 2015) (2015)
18. Gour, N., Khanna, P.: Multi-class multi-label ophthalmological disease 29. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recog-
detection using transfer learning based convolutional neural network. nition. In: Proceedings of the IEEE Conference on Computer Vision and
Biomed. Signal Process. Control 66, 102329 (2021) Pattern Recognition, pp. 770–778. IEEE, Piscataway, NJ (2016)
19. Camara, J., Rezende, R., Pires, I.M., Cunha, A.: Retinal glaucoma public 30. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the
datasets: What do we have and what is missing? J. Clin. Med. 11, 3850 inception architecture for computer vision. In: Proceedings of the IEEE
(2022) Conference on Computer Vision and Pattern Recognition, pp. 2818–2826.
20. Zhou, H.: Design of intelligent diagnosis and treatment system for oph- IEEE, Piscataway, NJ (2016)
thalmic diseases based on deep neural network model. Contrast Media 31. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.-C.:
Mol. Imaging 2022, 1–10 (2022) Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of
21. Rajee, R.J., Roomi, S.M.M., PooAnnamalai, V., Begam, M.P.: A transfer the IEEE Conference on Computer Vision and Pattern Recognition, pp.
learning approach for retinal disease classification. In: Proceedings of the 4510–4520. IEEE, Piscataway, NJ (2018)
2023 International Conference on Signal Processing, Computation, Elec- 32. Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.: Inception-v4, inception-
tronics, Power and Telecommunication (IConSCEPT), pp. 1–6. IEEE, resnet and the impact of residual connections on learning. In: Proc. AAAI
Piscataway, NJ (2023) Conf. Artif. Intell. 31(1), 11231 (2017)
22. Patil, Y., Shetty, A., Kale, Y., Patil, R., Sharma, S.: Multiple ocular dis- 33. Abdulrazzak, A.Y., Mohammed, S.L., Al-Naji, A.: NJN: A dataset for the
ease detection using novel ensemble models. Multimedia Tools Appl. 83, normal and jaundiced newborns. BioMedInformatics 3, 543–552 (2023)
11957–11975 (2023) 34. Zhu, Q.: On the performance of Matthews correlation coefficient (MCC)
23. Siddique, M.A.A., Ferdouse, J., Habib, M.T., Mia, M.J., Uddin, M.S.: Convo- for imbalanced dataset. Pattern Recognit. Lett. 136, 71–80 (2020)
lutional neural network modeling for eye disease recognition. Int. J. Online 35. Chicco, D., Jurman, G.: The advantages of the Matthews correlation coeffi-
Biomed. Eng. 18(9), 115–130 (2022) cient (MCC) over F1 score and accuracy in binary classification evaluation.
24. Amsalam, A.S., Al-Naji, A., Daeef, A.Y., Chahl, J.: Automatic facial palsy, BMC Genomics 21, 1–13 (2020)
age and gender detection using a raspberry Pi. BioMedInformatics 3, 455– 36. Vijayalakshmi, D., Nath, M.K.: A systematic approach for enhancement
466 (2023) of homogeneous background images using structural information. Graph.
25. Kasinski, A., Schmidt, A.: The architecture and performance of the face Models 130, 101206 (2023)
and eyes detection system based on the Haar cascade classifiers. Pattern
Anal. Appl. 13, 197–211 (2010)
26. Cuimei, L., Zhiliang, Q., Nan, J., Jianhua, W.: Human face detection algo-
rithm via Haar cascade classifier combined with three additional classifiers. How to cite this article: Al-Naji, A., Khalid, G.A.,
In: Proceedings of the 2017 13th IEEE International Conference on Mahmood, M.F., Chahl, J.: Computer vision for eye
Electronic Measurement & Instruments (ICEMI), pp. 483–487 (2017) diseases detection using pre-trained deep learning
27. Zoph, B., Vasudevan, V., Shlens, J., Le, Q.V.: Learning transferable architec-
techniques and raspberry Pi. J. Eng. 2024, e12410
tures for scalable image recognition. In: Proceedings of the Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition, pp. (2024). https://doi.org/10.1049/tje2.12410
8697–8710. IEEE, Piscataway, NJ (2018)