Available online at https://icsejournal.
com/index.
php/JCSE Journal of Computer Science an Engineering (JCSE) e-ISSN 2721-0251 Vol.
No.
February 2025, pp.
Employment of Convolutional Neural Networks in an Eye Disease Detection Application Leveraging Tensorflow.
Zainal Arifin1*.
Firman Santoso2.
Adi Susanto3 1,2,3Departmen Information Technology at University of Ibrahimy.
Situbondo.
Indonesian elibrahimy@gmail.
2 firman4bi@gmail.
3 adisusanto@ibrahimy.
* corresponding author
ARTICLE INFO
ABSTRACT
Article History:
Received June 9, 2024 Revised July 15, 2024 Accepted February 12, 2025 Vision impairment from cataract and glaucoma remains a leading global health challenge, with 2.
2 billion people affected worldwide and Indonesia registering the highest prevalence in Southeast Asia and second highest globally (Ministry of Health, 2017Ae2030 Roadmap.
WHO 2.
Early, accurate detection is essential to prevent irreversible blindness.
In this study, we develop and evaluate a native-architecture Convolutional Neural Network (CNN) to classify cataract and glaucoma using three novel non-fundus (AurealeyeA.
image subsets.
Trained for 100 epochs on a balanced dataset, our CNN 67 % accuracy, precision of 98.
8 %, and recall of 98.
5 % on a held-out test set.
The resulting TensorFlow-saved model is deployed in the browser via TensorFlow.
js, enabling real-time, platform-agnostic inference without specialized hardware.
This work demonstrates that non-fundus imaging, combined with lightweight CNN design and web deployment, can provide a practical, high-accuracy tool for early eye-disease screening in lowresource settings.
Keywords:
Blindness and Vision Impairment Convolutional Neural Network Machine Learning Classification Tensorflow Correspondence:
E-mail:
elibrahimy@gmail.
Introduction Vision impairment affects some 2.
2 billion people worldwide, with cataract and glaucoma as the two leading causes of preventable blindness.
In Indonesia alone, 3 % of the population suffers from vision impairment the highest rate in Southeast Asia and second highest globally driven chiefly by cataracts and, to a lesser extent, glaucoma.
Early, accurate screening is therefore critical to reduce the social and economic burden of vision loss.
Convolutional Neural Networks (CNN.
have shown great promise in medical image classification.
AlexNet achieved 98.
37 % accuracy on cataract versus normal fundus images but suffers from a large model size and slow inference .
, .
MobileNetV2 prioritizes speed and model compactness achieving 72 % accuracy yet sacrifices some classification performance .
Conversely, bespoke AunativeAy CNN architectures can be highly efficient on constrained hardware, but often at the cost of reduced accuracy and flexibility .
, .
, .
Moreover, most prior work relies on high-resolution fundus photography, requiring specialized equipment and limiting accessibility in low-resource settings.
Non-fundus (Aureal-eyeA.
imaging via standard digital or smartphone cameras remains underexplored despite its potential for widespread screening .
, .
, .
To bridge this gap, we propose a lightweight, native-architecture CNN that classifies three nonfundus .
eal-ey.
image classesAicataract, glaucoma, and normalAicollected via standard digital Our pipeline resizes each image to 100y100 pixels, applies repeated convolution and MaxPooling layers, and uses ReLU and Softmax activations over 100 training epochs, achieving 67 % accuracy on a balanced 3 000-image dataset.
Crucially, we export the trained model to TensorFlow.
js, enabling in-browser inference without dedicated GPU hardware.
Journal of Computer Science an Engineering (JCSE) Vol.
No.
February 2025, pp.
e-ISSN 2721-0251 Method The methodology of this research is structured into three distinct components: Research Method.
Data Collection Method, and Data Analysis Method.
1 Research Method The research methodology adopted in this study is an experimental method with an objective quantitative approach, involving the development and evaluation of a Convolutional Neural Network (CNN) model.
2 Data Collection Method A comprehensive literature review served as the primary data collection method in this research, with the aim of tackling eye disease-related challenges and implementing deep learning techniques using classification methods to identify the optimal model based on training outcomes.
The literature review encompassed a wide range of sources, including academic journals, electronic books, and relevant websites.
3 Data Analysis Method Data analysis is employed to tackle the research problem by utilizing the gathered data and The data analysis methodology comprises the following stages:
Figure 1.
Workflow of Data Analysis 1 Datasets The data collection process involved gathering a total of 3000 images in .
jpg format.
These images comprised cataract, glaucoma, and normal eye images, acquired on February 28, 2024, from the Roboflow.
com platform.
Figure 2.
Normal and Cataract Eye Dataset 2 Preprocessing Preprocessing is an essential phase in data preparation, aimed at addressing potential problems that could impair the effectiveness of data processing.
This stage encompasses various techniques, including rescaling, shearing, zooming, and horizontal flipping.
3 Classification Classification encompasses the convolution stage, utilizing image input.
The convolution process involves applying filter values to a matrix, considering the three red, green, and blue channels of image pixels.
Journal of Computer Science an Engineering (JCSE) Vol.
No.
February 2025, pp.
e-ISSN 2721-0251 Figure 3.
Image Pixel Values and Kernel Test Matrix Figure 3 depicts the computational procedure for each channel .
ed, green, and blu.
involving multiplication with a 3x3 kernel or filter.
This iterative process involves shifting the kernel by 1 stride in each channel, generating the calculated values for each channel.
The image pixel values are extracted using Google Colab to obtain the pixel values for each channel separately.
Figure 4.
Convolutional Calculation with Filter 3x3 The convolution calculation is carried out on each channel matrix by multiplying a 3x3 matrix element-wise according to its rows and columns, followed by summation to produce a result value This process is executed with a stride of 1 from left to right.
Figure 5.
Activation of RGB Convolutional Values with ReLU Following the convolution computation, the values from each channel are aggregated to yield the total RGB convolution value.
Subsequently.
ReLU activation is applied.
During ReLU activation, all negative values are transformed into 0, producing the convolution outcome for figure 5.
Figure 6.
Pooling Layer The max-pooling stage employs a 2x2 kernel and selects the maximum value at each stride position.
This max-pooling operation utilizes a stride of 2, resulting in the output being fed into the dropout The dropout phase serves to reduce the complexity of the constructed model by eliminating unused neural networks.
Journal of Computer Science an Engineering (JCSE) Vol.
No.
February 2025, pp.
e-ISSN 2721-0251 Figure 7.
Flatten Layer The flatten layer serves to convert a multidimensional matrix into a one-dimensional matrix or vector.
This operation transforms the output of the 2x2x32 dropout layer into a new vector, as illustrated in Figure 7.
Figure 8.
Illustration of Fully Connected Layers.
The following equation .
, .
, .
is the hidden layer formula .
, .
The output values obtained from the flatten operation will be processed by the fully connected layer, utilizing computations from each of its hidden layers.
R1 = .
*0,.
*0,.
*0,.
*0,.
= 112,3
R2 = .
*0,.
*0,.
*0,.
*0,.
= 449,2
R3 = .
*0,.
*0,.
*0,.
*0,.
= 561,5
R4 = .
*0,.
*0,.
*0,.
*0,.
= 336,9
The first hidden layer undergoes multiplication with distinct weight values, yielding the values from the preceding computation.
Subsequently, the results of this hidden layer calculation are further multiplied using different weight values to produce the values S1.
S2.
S3.
S4 which constitute the next hidden layer.
S1 = .
,3*0,.
,2*0,.
,5*0,.
,9*0,.
= 437,97
S2 = .
,3*0,.
,2*0,.
,5*0,.
,9*0,.
= 729,95
S3 = .
,3*0,.
,2*0,.
,5*0,.
,9*0,.
= 291,98
Journal of Computer Science an Engineering (JCSE) Vol.
No.
February 2025, pp.
e-ISSN 2721-0251 S4 = .
,3*0,.
,2*0,.
,5*0,.
,9*0,.
= 583,96
Every neuron in S1.
S2.
S3.
S4 will undergo multiplication with distinct weights once more, leading to the generation of the values Cataract.
Glaucoma, and Normal.
Cataract = .
,97*0,.
,95*0,.
,98*0,.
,96*0,.
= 1021,93
Glaucoma = .
,97*0,.
,95*0,.
,98*0,.
,96*0,.
= 613,158
Normal = .
,97*0,.
,95*0,.
,98*0,.
,96*0,.
= 204,386
The subsequent stage entails computing the softmax activation employing the exponential formula:
Cataract divided by the sum of cataract glaucoma normal.
This procedure is executed thrice, once each for cataract, glaucoma, and normal.
s(O.
s(O.
s(O.
= 1
0,5 0,3 0,1 = 1
Results and Discussion Subsequent to achieving uniform image sizes, the image data distribution will advance to the classification stage.
This stage in the research utilizes several parameters outlined in Table 1.
Table 1.
Hyperparameters Employed Image Dimensions Params Epoch 100 x 100px Batch Size Optimizer RMSprop Utilizing the parameters employed in conjunction with the Native CNN architecture and the implemented model, the outcomes are depicted in Figure 9.
Journal of Computer Science an Engineering (JCSE) Vol.
No.
February 2025, pp.
e-ISSN 2721-0251 Figure 9.
Accuracy and Loss Trajectory Utilizing Convolutional Neural Network (CNN) for eye disease classification, a remarkable accuracy 67% is attained, as illustrated in Figure 9.
The following equation .
is the exponential formula for softmax activation.
Consequently, the weight value corresponding to the probability of cataract emerges as the largest, at 0.
5, implying that the input image depicts a cataract.
Figure 10.
Employing TensorBoard for Model Monitoring An evaluation of the classification algorithm was conducted using a sample dataset of 10 instances.
The results demonstrated a classification testing accuracy of 98.
The saved model will be deployed as a web application utilizing TensorFlow.
Conclusion This study investigates eye disease classification using the Convolutional Neural Network (CNN) method with a native CNN architecture.
The study's novel contribution lies in the utilization of three non-fundus image classes: cataract, glaucoma, and normal.
The preprocessing stage involves resizing images to 100x100 pixels.
The CNN model is implemented using 100 epochs for its hyperparameters.
The achieved classification accuracy is 98.
References