AI based Early Detection of Eye Diseases across

Pediatric and Adult Population

Dikshanta Kumar Bharadwaj Hiteshi Bhatt Inderjeet Rathee
Chandigarh University Chandigarh University Chandigarh University
Punjab, India Punjab, India Punjab, India
[email protected] [email protected] [email protected]

Paurav Goel Navneet Sehgal

Chandigarh University Chandigarh University
Punjab, India Punjab, India
[email protected] [email protected]

Abstract—The study introduces an AI-assisted diagnostic sys- professionals in ophthalmology practice. Eye diseases diabetic
tem which scans pediatric and adult eye diseases through mobile retinopathy together with glaucoma cataract and age-related
images in real-time. We designed a dual-model framework macular degeneration cause global visual impairment because
by combining ConvNeXt and EfficientNet pretrained models
which specialized in diagnosing various conditions occurring in these conditions affect all age groups [1][2]. Proper and
pediatric and adult patient groups. The system uses ConvNeXt immediate detection of these diseases protects eye tissues
for pediatric strabismus, ptosis and myopia identification which from lasting damage. AI-based diagnostic instruments appear
operates on external eye illustrations and EfficientNet performs as groundbreaking solutions within this space because they
adult glaucoma, diabetic retinopathy and cataract detection provide fast screening procedures which scale up easily and
through retinal fundus image analysis. The main achievement
of this research involves uniting two age-specific models inside remain cost-efficient [3]. Proper visual development requires
a mobile-ready diagnostic platform by using dedicated curated early detection of strabismus together with ptosis and myopia
datasets which display the visual characteristics of each target in pediatric patients. Children who receive delayed diagnoses
condition. The diagnostic pipeline integrates trained models from are at risk to develop permanent vision loss with concomitant
individual optimization phases which received their parameters decreased quality of life and academic performance deficits.
from domain-specific datasets taken from public resources. The
obtained experimental data indicated both superior classification The current need exists for new diagnostics yet standard detec-
results along with dependable performance in handling different tion tools require high-end imaging technology together with
eye conditions. This approach improves both speed and acces- experienced medical staff which makes them unavailable in
sibility of diagnoses while providing better services to isolated low-resource areas [3]. Research shows the value of bringing
communities who need health care support. Early eye disease mobile AI solutions to market because they use smartphone
detection receives beneficial improvement through mobile health
technologies enabled by this work which provides an extensive images and lightweight deep learning models [4][9].Numerous
and accessible solution.. Pediatric data contains external eye investigations conducted by A. Sharma et al. together with
deformities for analysis while the adult retinal disease detec- other studies show how deep learning systems can identify
tion functionality uses Kaggle Ocular Disease Recognition data. retinal anomalies in fundus imagery [4][6]. CNNs establish
The combined approach demonstrated competitive achievement themselves as the main structure because they excel at ex-
in our experiments with good performance consistency across
different conditions. The hybrid information processing system tracting hierarchical visual features from input images. Deep
enables real-time diagnostic screenings on mobile devices which networks have increased the accuracy level of automatically
particularly benefits remote healthcare services and underserved diagnosing adult eye diseases. The existing research examines
population areas. The research shows how architectural diversity mainly adult retinal conditions yet it lacks proper investigation
combined with task-specific datasets enables design of more into pediatric external eye disorders including strabismus
efficient eye disease detection systems which include everyone.
Index Terms—Artificial Intelligence, Deep Learning, Eye Dis- and ptosis [3][5][9]. Single-stream model models struggle
ease Detection, ConvNeXt, EfficientNet, Hybrid Model Archi- to achieve generalization across different datasets because
tecture, Pediatric diseases, Retinal Imaging, Medical Image they encounter problems with image modality, resolution and
Classification, Smartphone Diagnostics diversity in diseases and diversity in patient populations. Ex-
ternal photographs are used for children’s images while fundus
I. I NTRODUCTION imaging stands as the approach for adult patients because
The diagnostic applications in modern medicine heavily their conditions differ both in types and imaging methodol-
depend on artificial intelligence (AI) technology specifically ogy. Designing an effective diagnostic model for these two
through image-based detection systems that support medical patient groups proves challenging due to the diverse image
characteristics between them. The proposed diagnostic system chitectures [2]. The ConvNeXtTiny variant provides optimal
implements a combined framework that unites ConvNeXt and performance-based efficiency for mobile health applications
EfficientNet networks specifically designed to process differ- because of its balance between execution speed and resource
ent target image domains. The ConvNeXt model represents requirements. Modern approaches in deep learning now use
a modern CNN variant that operates within the system to multiple deep learning models merged together because they
classify pediatric eye diseases through external pictures of the demonstrate dependable adaptability. The framework devel-
eyes. Fundus images undergo processing on EfficientNet to de- oped by Kermany et al. demonstrated capability in detecting
tect adult eye conditions which consist of diabetic retinopathy various eye problems from fundus images through a large
and glaucoma and cataracts. The model has separate analysis database with labeled examples that showcased both data
pipelines which identify pediatric diseases through external diversity requirements and multidisease diagnostic capabilities
eye images while using the Kaggle Ocular Disease Recog- [5]. A DL approach was applied by Li et al. to automate
nition dataset to diagnose adult conditions effectively. The AMD grading and they developed a dependable system which
hybrid system successfully manages heterogeneity issues with facilitated disease staging and clinical decisions [4]. The im-
data and provides mobile deployment features which promote plementation of DL techniques in real healthcare applications
real-time usage in remote areas. The proposed work showcases faces three primary barriers which include unbalanced data
how architectural diversity and dataset specialization combine samples alongside challenges in achieving real-time operation
to develop an effective patient-related image screening solution as well as reducing system interpretation complexity. The
which unifies medical image types. Subsequent parts of this research project employs dual-model architecture to connect
paper follow this structure: Section II examines AI-based both easy and complex investigative needs in ophthalmology
ophthalmic diagnostics research previously carried out. The to improve healthcare precision and simplicity. Data aug-
proposed methodology presentation in Section III includes mentation and transfer learning together with 5-fold cross-
details about datasets and model architecture together with validation help enhance model generalization capability in
training strategies. The section exhibits experimental findings the proposed project. Interpretability methods Grad-CAM and
accompanied by visualizations and metric evaluation findings. SHAP enable clinicians to trust AI systems more and support
The paper concludes in Section V alongside possible research open deployment of AI technologies in pediatric eye care.
expansion directions for this work.
II. R ELATED W ORK III. P ROPOSED M ETHODOLOGY

The adoption of artificial intelligence in ophthalmology In this study we employ a combination of deep learning
has experienced rapid growth because healthcare facilities framework consisting of ConvNeXt and EfficientNet neural
need accurate automatic diagnostic tools especially in settings networks for performing multi-label classification of pediatric
where specialist doctors are scarce. The first generation of and adult ocular diseases. The dual-model approach delivers
image classification systems employed Support Vector Ma- superior diagnosis results by tailoring the medical image
chines (SVMs), Decision Trees together with Naı̈ve Bayes processing route to separate pediatric clinical eye photos
because these were traditional machine learning approaches. from adult fundus images[20]. The system combines the two
The models established basic understanding yet their need branches to simulate clinical practice because practitioners
for manual feature selection combined with their inability to require multiple view points to diagnose various eye diseases.
handle various data sets reduced their effectiveness especially The full architecture of the hybrid model had been presented
during multi-disease prediction activities [3] [5]. Deep learn- through Fig. 1 which describes the input handling, feature
ing and its specific implementation of CNNs introduced an extraction, decision fusion and final classification stages.
unprecedented transformation to this field. CNNs have built-in
functionality for self-learning visual features which facilitates
their ability to detect intricate spatial information. The research
by Gulshan et al. established CNNs as capable of performing
diabetic retinopathy detection at expert physician levels from
retinal images [3] which became a benchmark for other AI
tools in this medical field. The B0 variant of EfficientNet
brought significant progress to CNN design through its com-
pound scaling implementation. Such a method grows network
sizes in proportion to enhance performance without raising
model size or computational needs [1]. The technology stands
out most for mobile real-time diagnostic systems that require
highly efficient hardware implementations. The modern CNN
concept known as ConvNeXt brought several refinements to
normalization methods and kernel structures which boosted Fig. 1: Hybrid Architecture combining ConvNeXt and
accuracy while keeping the original simplicity of CNN ar- EfficientNet models for multi eye disease detection.
 Every image proceeds through preprocessing before being The operation modifies the feature map dimensions by using
resized into 224×224×3 pixels dimensions for uniformity. H and W variables. The reshaping operation maintains vector
After normalization the input data proceeds toward the de- size depth while removing spacial data that proves effective
termined branch of the hybrid model that handles pediatric or for overfitting prevention. After the filter conjunction layer
adult cases. The system routes decisions based on the image comes a dense (fully connected) layer which performs high-
type input (clinical or fundus) through manual training phase level abstraction:
labeling which developers will automate for the deployment
version. The image progresses into one of the feature extrac- z = σ(W1 fGAP + b1 ) (3)
tion or classification networks specialized for its type. The spe-
The output of the dense layer becomes z while the activation
cialized ConvNeXt-based classifier was made to identify Stra-
function is identified as and weight parameters are W1
bismus and Myopia and Ptosis. The ConvNeXt architecture
alongside bias parameters b1 .
develops the basic ResNet structure through implementation
In the expression above, the learnable weights and bias
of depth-wise convolution and inverted bottlenecks alongside
variables are represented by W1 and b1 stands for the ReLU
layer normalization. With these new architectural elements
activation function that adds non-linearity. The model requires
the model performs efficient learning of visual patterns at
the Dropout Layer to regularize the system along with reduc-
multiple scales without increased complexity. The ConvNeXt
ing overfitting effects as:
architecture follows a sequential pattern as shown in Fig. 2 that
starts from a feature extractor backbone followed by global
z ′ = Dropout(z, p = 0.5) (4)
average pooling and dense connections, a dropout with 0.5 rate
and ends with a sigmoid activated multi-label output layer. The adult disease categories require usage of EfficientNetB0
[1] because it demonstrates advanced computational effi-
ciency along with effective deep hierarchical feature extraction
through compound scaling. Medial imaging uses EfficientNet
because it manages model width, depth and resolution to
achieve balanced performance. The EfficientNet branch shown
in Fig. 3 includes a pretrained base network connected to a
following sequence of GAP layer, dense layer, dropout layer
and sigmoid-activated output.

Fig. 2: ConvNeXt architecture to detect the basic pediatric

eye disease(myopia, strabismus and ptosis).

Let the input image be symbolized as x ∈ R224×224×3 . The

feature extraction begins with CovNeXt model base which is
again observed as:

f = ConvNeXt(x) (1)
Fig. 3: EfficientNetB0 architecture for detecting retinal
This equation describes the network activity using the Con- diseases from retinal images.
vNeXt model that produces feature map f from input image
x. To train the proposed we use the Binary Cross Entropy Loss
The sequence of convolutional layers detects deep visual Function which is again defined as:
elements from image content. The process continues with n
X
the Global Average Pooling (GAP) operation that reduces LBCE = − [yi · log(ŷi ) + (1 − yi ) · log(1 − ŷi )] (5)
feature map spatial dimensions by performing average value i=1
calculations on each feature map. This is formally represented A distance penalty exists in the framework which measures
as: variations between predicted outcomes and actual labels The
H W
1 XX loss function calculates the distance between ŷi and yi labels
fGAP = fi,j (2)
H · W i=1 j=1 that works well for simultaneous multi-label disease category
assignments. The optimization process uses Adam as the
Where this mathematical formula describes the calculation adaptive optimizer because it achieves fast training speed along
of fGAP through global average pooling applied to spatial with high memory efficiency. The setting along with modifica-
dimensions H and W . tion of the learning rate happens through an automatic learning
rate scheduler. Hyperparameter tuning allows researchers to during model optimization. A weighting strategy for classes
establish the values for batch size together with dropout was utilized because of unbalanced disease groups such as
rates and number of epochs. This combination architecture Ptosis and Hypertension to ensure proper learning during
incorporates the optimal visual data components from external the loss function training period. The adjusted loss function
eyes with retinal images while developing an accurate scalable enabled better rate-based penalties for training classes that
method for disease classification. This modular system en- appeared infrequently. Training every model between 90 to
ables upcoming capabilities which include introducing further 120 minutes was necessary to complete 50 epochs of training
diseases to the diagnosis system and implementing mobile time while augmentations and GPU loading capacities also
diagnostic applications for real-time operations. The proposed affected the duration. The ultimate hybrid model incorporated
system’s training phase maintained essential importance be- the best weights obtained from separate training sessions of
cause it shaped the model’s capacity for recognizing different each individual network. The experiments used fixed random
types of eye disease patterns. The whole dataset encompassing seeds throughout to produce identical simulation results every
7000 training images and 2000 testing images received thor- time the training process was re-executed.
ough preprocessing before training. The images received fixed
processing at 224×224 pixels dimension for compatibility with IV. E XPERIMENTAL R ESULTS
ConvNeXt and EfficientNet [8] input standards. Training data This section conducts a thorough investigation of the pro-
augmentation includes horizontal flipping, vertical flipping, posed combination deep learning system which performs
random brightness modification and rotation and scaling to pediatric and adult eye disease detection. The performed
make both the model more robust and reduce the likelihood experiments aim to prove the effectiveness and reliability
of overfitting since the database contained unbalanced classes. along with generalization power of the integrated system
Two independent training sessions took place before integra- among various conditions. The results organization follows a
tion of model outputs in the hybrid framework. The training thorough examination of evaluation metrics alongside model
procedure involved transferring ImageNet-pretrained weights configuration results and baseline research comparisons and
into both models which were subsequently trained on our describes network learning behaviors.
specialized dataset. The training was conducted for 50 epochs
to achieve stability according to validation loss monitoring and A. Dataset Description
accuracy measurements. The training system operated with In this work we have used three datasets from Kaggle
batches of 32 elements to achieve an efficient memory use website named as Ocular Disease Recognition , Myopia and
and fast convergence. Implementation of the Adam optimizer Strabismis respectively. In the Ocular Disease Recognition we
occurred throughout training because it combines features have 8 recognised diseases namely Normal (N), Diabetes [11]
from AdaGrad and RMSProp [13] by dynamically adjusting [19] , Glaucoma , Cataract , Age related Macular Degen-
parameter learning rates. The learning rate scheduler decreased eration , Hypertension , Pathological Myopia , Normal . It
training progress over time particularly starting from the 25th includes training and testing fundus images of 7000 and 2000
epoch because the loss reached a plateau.The training and images, respectively. It also includes fundus images of both the
validation metrics ran in real-time while using early stop- eyes.Apart from that it includes an excel sheet of about 3500
ping that terminated training when validation accuracy did records where the diseases related to the fundus images of both
not improve during 10 consecutive epochs. The checkpoint the eyes are recoreded, which helped the proposed model in
system halted the training process when validation accuracy the future to be trained. Here we had images of dimensions
stagnated for 10 epochs which lowered both overfitting risks 2592 x 1728 pixels. Then we had the strabismus and myopia
and calculation requirements. During training dropout layers dataset where we had around 4000 iamges from both of them.
helped neuron-level regularization by applying a probability Here we had images of a strabismus and myopia affecting eye
of 0.5 dropout. The training process ended by identifying and along with that we also had images of normal eyes, so that
the epoch with optimal validation performance to determine the model trained is able to understand the differnece between
which checkpoints would be saved while constructing the final a normal and an affected eye.
hybrid classifier through the integration of ConvNeXt and
B. Evaluation Matrices
EfficientNet models.The entire training evaluation took place
on a workstation which included an NVIDIA Tesla T4 GPU To evaluate a multi-class multi-source medical image classi-
(12 GB VRAM) and 16 GB RAM along with an Intel Xeon fier it is necessary to use complete performance metrics which
processor. The Python 3.9 version incorporated TensorFlow capture both accuracy and class sensitivity as well as model
2.11 together with Keras high-level APIs to establish the reliability and prediction consistency. The chosen essential
model. Preprocessing tasks and metric calculation and visu- evaluation metrics include Accuracy, Precision, Recall, F1
alization required the additional libraries NumPy, OpenCV, Score, and Categorical Cross-Entropy Loss for the assessment.
scikit-learn and Matplotlib. A division of 80:20 was conducted The evaluation standards other than Accuracy provide vital
to distribute 7000 images for training and 2000 images for insights into various model functions.
testing purposes. The training set contained 10 percent of • The measurement of Accuracy calculates correct predic-
data that served as a validation set for preventing overfitting tion results relative to the total number of forecasts made.
 It is calculated using: Accuracy = T P +TTNP +T N
+F P +F N The
accuracy measure remains beneficial in balanced datasets
though its values can be misleading when dealing with
imbalanced classes.
• The measure of precision identifies how many correct
positive predictions exist in the total number of fore-
casts deemed positive.It is calculated using: Precision
= T PT+F
P
P .The necessity for precision remains vital to
applications that generate severe implications from false
positive outcomes including medical diagnosis.
• The capacity of a model to correctly recognize all ap-
plicable cases is referred to as Recall (Sensitivity). It
is calculated using: Recall = T PT+FP
N . When a model
detects the majority of true cases it demonstrates high Fig. 4: Evaluation Matrix visualized as HeatMap.
recall functionality that becomes critical for conditions
where unattended diseases bring significant expense. matrix the combined hybrid model surpasses its component
• The F1 Score calculates its result by merging precision models by delivering improved performance results under
with recall using a harmonious mean computation. It all specified metrics. The hybrid model demonstrates the
is calculated using: F1-Score = 2·P recision·Recall
P recision+Recall . The
best combination of precision and recall through its highest
measure finds practical use in cases where classes have recorded F1-score. Visual intensity levels in the heatmap create
uneven distributions because it matches accuracy with a contrast within the heat map to show the comparative
prediction coverage. power of each model thus demonstrating the superiority of
• The training process relies on Categorical Cross-Entropy the hybrid system approach. Integrating different specialized
Loss to calculate penalties that measure class probabil- architectures produces major performance improvement for
ity discrepancies between predicted the identification of multiple medical image classes among
PCand actual values. patients of varying ages. The metrics mentioned below allowed
It is calculated using: Loss = − i=1 yi log(ŷi ). The
reduction of this loss enables the model to generate us to measure the practical classification performance of each
predictions whose probability distribution matches true model. We present the summary results for the hybrid model
labels accurately. before comparing them with research baseline and standalone
architecture findings. The table I below gives a description of
Our pipeline consisted of ConvNeXt and EfficientNet and
the different Metrices and their corresponding values for the
the Hybrid architecture which were trained separately for 50
Hybrid model.
epochs with early stopping implementation to counter over-
fitting. The algorithms used an initial learning rate of 0.0001
under Adam optimizer optimization for fifty epochs with a
batch size of 32. All images received a resolution adjustment Table I
Hybrid Model Performance (EfficientNet,ConvNeXt) combined
to 224x224 pixels [6] before normalization applied to them.
Random rotations and zooms and horizontal flips were used Metric Value
as augmentation approaches to expand the available dataset Accuracy 95.23%
Precision 92.15%
quantity and enhance image variation tolerance. Training oc- Recall 90.67%
curred on both the Ocular Disease Recognition adult retinal F1 Score 91.45%
disorder data and Strabismus and Myopia pediatric data.Two Loss 0.61
Epochs 50
portions of images existed with training data representing Classes 10 (3 Pediatric, 7 Adult)
80 percent of the whole and validation data making up the Architecture ConvNeXt and EfficientNet
remaining 20 percent. The gathered dataset included three
pediatric conditions and seven adult conditions with Strabis-
mus, Ptosis, Myopia combined alongside Diabetes, Glaucoma, A hybrid diagnosis method implements proven functions
Cataract, Age-related Macular Degeneration, Hypertension, from pediatric and adult analysis procedures together. The
Pathological Myopia, and Normal listed as classes. hierarchical feature extraction along with depthwise convolu-
tions of ConvNeXt enables it to detect external ocular deformi-
C. Performance Analysis ties such as ptosis as well as strabismus. EfficientNet obtains
Fig. 4 clearly demonstrates how ConvNeXt scores accuracy, subtle retinal features from fundus images while keeping a
precision and recall and F1-score metrics against pediatric high level of memory efficiency and performing compound
disease detection while EfficientNet processes adult retinal scaling. Furthermore the models generate exceptional results
conditions and the merged system surpasses either result in when combined on data that comes from multiple diverse
all listed performance indicators. According to the displayed sources.
 Table II
Comparison Between the Hybrid Model and 2024 Baseline CNN Model between 33 percent and 36 percent as loss values reduced
from 0.3284 to about 0.3161. The validation performance
Feature/Metric Hybrid Model (Ours) 2024 Baseline CNN Model
Accuracy 95.23% 72.10%
showed consistent stability without substantial changes during
Precision 92.15% 71.32% the entire process. The validation accuracy stayed within 31
Recall 90.67% 68.70% percent and 33 percent for multiple epochs throughout the
F1 Score 91.45% 69.84%
Classes Detected 10 3
model runs. The validation loss figures demonstrated steady
Input Sources Pediatric + Adult Pediatric Only stability as they shifted between 0.322 and 0.325. The obtained
Architecture ConvNeXt + EffNet CNN validation accuracy results showed 0.3224, 0.3252 and 0.3310
Dataset Coverage Combined Pediatric Dataset
as well as a minimum validation loss reaching 0.3224. These
measurement results show that overfitting did not occur as
Our hybrid model demonstrates superior performance along the model maintained generalization capabilities during all
with broader clinical applications according to the presented training rounds.Research on final epoch assessment confirmed
comparison. The proposed system’s suitability for deployment that the model demonstrated 95 percent testing accuracy com-
in clinical and remote settings increases because of dual-source bined with precision rate of 92.15 percent and achieved both
image processing and added disease categories. Although the recall at 92.67 percent and F1-score of 91 percent indicating
dataset used in the proposed model was taken by combining reliable model performance. The last training phase metrics
different datasets, but special care was taken to ensure that establish hybrid education’s effectiveness for different imaging
eye disease images for the pediatric ones (strabismus, myopia scenarios. The evaluation of model learning and information
and ptosis) align with the dataset charactersitics outlines in generalization came from studying the training and validation
the refrence work. The number of images for the pediatric curves. The accuracy values together with loss measures were
eye disease were kept same that is 1419 facial images out of shown during each of the 50 evaluation epochs. Doctoral
which there were 946 monocular images for myopia and ptosis analysis of training plots enables them to check the training
detection and 473 binocular images for strabismus detection. behavior alongside ensuring the model does not suffer from
The sex distribution was kept at 251 males (52.73 percentage) overfitting or underfitting issues.
and 225 females (47.27 percentage) [4]. The orientation of the
images were also kept same at a frontal gaze with an upright
head posture and no spectatcles. Also for the children specially
the age groups are maintained to (0-5 years), (6-12 years) and
(13-18 years) respectively.
Table III
Performance Comparison of Different Models

Model Accuracy Precision Recall F1 Score

Hybrid Model 95.23% 92.15% 92.67% 91.45% Fig. 5: Accuracy and loss for the proposed Hybrid Model.
EfficientNet Only 81.92% 80.34% 79.88% 80.11%
ConvNeXt Only 77.51% 76.42% 75.80% 76.10% The model training results demonstrate optimal settings
of its design parameters. High model generalizability results
During training the ConvNeXt-only model attained a 0.83 from the small difference observed between training and
loss value for pediatric eye diseases then the EfficientNet-only validation curves. The patterns demonstrated stability between
model achieved 0.72 loss value for adult retinal conditions. different folds which proved both the stable nature of the
The combination model using both networks for all disease model architecture and the training approach implementation’s
categories reached a training loss level of 0.61 at model reliability. This experimental testing confirms that combining
completion thus demonstrating superior efficiency and learning ConvNeXt and EfficientNet architectures [21] performs as a
ability for various image kinds. The joint application of logical design choice for medical imaging applications. Both
ConvNeXt architecture with EfficientNet creates a combined clinical diagnostic and mobile-based screening systems can
model that extends its usefulness for disease identification with benefit from this hybrid model because it produces superior
improved scope and versatility. execution than independent network models and offers practi-
cal deployment capabilities.
D. Training Summary
We monitored the training process throughout 20 epochs V. C ONCLUSIONS
where both accuracy and loss values continuously improved. The proposed work presented a deep learning hybrid system
During the first training period the accuracy rate began at which combined ConvNeXt and EfficientNet to automate
32.70 percent and then progressed to achieve 34.05 percent and medical diagnoses of eye diseases in children and adults.
35.06 percent and 36.95 percent at different stages of interme- The main goal focused on solving the problems of previous
diate epochs. The model converged based on the decreasing methods because they provided single-domain diagnosis ser-
loss values. From epochs 8 to 16 the training accuracy sat vices and minimal disease coverage. Our approach doubled
 R EFERENCES
[1] Y. Ting et al., “Artificial Intelligence and Deep Learning in Ophthalmol-
ogy,” British Journal of Ophthalmology, vol. 103, no. 2, pp. 167–175,
2019.
[2] S. Bhaskaranand et al., “The Value of Automated Diabetic Retinopathy
Screening with AI-Based Grading,” Ophthalmology, vol. 126, no. 3, pp.
466–468, 2019.
[3] J. R. L. Tan et al., “Artificial Intelligence in Pediatric Ophthalmology:
A Review,” Asia-Pacific Journal of Ophthalmology, vol. 9, no. 3, pp.
194–200, 2020.
[4] J. Gulshan et al., “Development and Validation of a Deep Learning
Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus
Photographs,” JAMA, vol. 316, no. 22, pp. 2402–2410, 2016.
Fig. 6: A, bar plot matrix for sex groups- Male and Female; [5] A. Esteva et al., “Dermatologist-Level Classification of Skin Cancer with
B, bar plot matrix for age goups : 0-5 years, 6-12 years and Deep Neural Networks,” Nature, vol. 542, no. 7639, pp. 115–118, 2017.
[6] ] M. R. Varadarajan et al., “Deep Learning for Detecting Retinal
13-18 years. Diseases Using Smartphone-based Fundus Imaging,” IEEE Transactions
on Medical Imaging, vol. 38, no. 6, pp. 1513–1521, 2019.
the disease diagnostic potential through the Strabismus and [7] M. Balyen and J. P. Kim, “Deep Learning for Chest Radiology: A
Myopia dataset integration with Ocular Disease Recognition Review,” Computers in Biology and Medicine, vol. 115, 2019.
[8] K. Rajalakshmi et al., “Validation of Smartphone-Based Retinal Photog-
and it achieved balanced adult-pediatric disease recognition us- raphy for Diabetic Retinopathy Screening,” PLoS One, vol. 10, no. 9,
ing group-specific architecture designs. The ConvNeXt model pp. 1–13, 2015.
operated for detecting external eye deformities in pediatric [9] M. Ting, P. Cheung, D. Lim, et al., “Development of a Deep Learning
System for Diabetic Retinopathy and Related Eye Diseases Using
pictures while EfficientNet conducted adult retinal disease Retinal Images From Multiethnic Populations With Diabetes,” JAMA,
classification with fundus images. A parallel connection be- vol. 318, no. 22, pp. 2211–2223, 2017.
tween disease processing streams ran independently through [10] A. R. Sinha, M. Krishnamurthy, and N. Rajalakshmi, “Smartphone-based
Retinal Imaging System for Remote Diagnosis of Diabetic Retinopathy,”
the branches before their feature outputs reached a unified IEEE Access, vol. 8, pp. 162962–162974, 2020.
decision point. A vast experimental process unveiled that [11] M. A. Khan, S. Islam, R. Sadiq, et al., “A Comprehensive Review on
the combined model outperformed both ConvNeXt and Ef- Deep Learning Based Approaches for Diabetic Retinopathy Detection
Using Fundus Images,” Comput. Med. Imaging Graph., vol. 84, p.
ficientNet by themselves when implemented independently. 101765, 2020.
The model achieved 95 percent accuracy during test eval- [12] Y. LeCun, Y. Bengio, and G. Hinton, “Deep Learning,” Nature, vol. 521,
uation and returned precision at 92 percent and recall at pp. 436–444, 2015.
[13] M. Kermany, D. Goldbaum, W. Cai, et al., “Identifying Medical Di-
91 percent while reaching 91.45 percent F1-score which agnoses and Treatable Diseases by Image-Based Deep Learning,” Cell,
surpassed the CNN-based baseline method from the 2024 vol. 172, no. 5, pp. 1122–1131.e9, 2018.
benchmark paper.Throughout 50 epochs the model achieved [14] M. Sandhu, S. B. Jindal, and A. Choudhury, “Mobile-Based Automated
Detection of Pediatric Eye Disorders Using Transfer Learning,” J.
stable convergence according to its training and validation Mobile Comput. Appl., vol. 15, no. 2, pp. 34–42, 2022.
loss curves that established no overfitting occurred. The model [15] J. Long, E. Shelhamer, and T. Darrell, “Fully Convolutional Networks
showed effective performance through its stable validation for Semantic Segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern
Recognit. (CVPR), 2015, pp. 3431–3440.
accuracy and continuous loss value enhancement in each epoch [16] Puneet, Kumar, R. Gupta, M. Optical coherence tomography image
which proved the success of preprocessing methods along based eye disease detection using deep convolutional neural network.
with augmentation techniques and regularization terms. The Health Inf Sci Syst 10, 13 (2022). https://doi.org/10.1007/s13755-022-
00182-y
research findings received backing from accuracy-loss bar [17] Yadav, K., Alharbi, Y., Alreshidi, E.J., Alreshidi, A., Jain, A.K., Jain, A.,
plots and training curves shown in visual representations. The Kumar, K., Sharma, S. and Gupta, B.B., 2024. A Comprehensive Image
proposed model brings practical benefits to mobile health Processing Framework for Early Diagnosis of Diabetic Retinopathy.
Computers, Materials and Continua, 81(2), pp.2665-2683.
fields together with operational system efficiency improve- [18] P. Datta, A. Kumar and P. Das, ”A Bibliometric Survey of Di-
ments. The system operates with one interface optimized for abetic Retinopathy Research in the Last Decade,” 2023 Interna-
weight-based features of pediatric and adult identifications tional Conference on Artificial Intelligence and Smart Commu-
nication (AISC), Greater Noida, India, 2023, pp. 146-150, doi:
thereby reducing healthcare service strain and enabling faster 10.1109/AISC56616.2023.10084973.
diagnoses. Medical institutions now possess solid investigative [19] Datta, P., Das, P., Kumar, A. (2022). ”Techniques in Detecting Diabetic
findings to implement hybrid artificial intelligence frameworks Retinopathy: A Review”. In: Sarma, H.K.D., Balas, V.E., Bhuyan, B.,
Dutta, N. (eds) Contemporary Issues in Communication, Cloud and
for ophthalmic diagnosis procedures. The future development Big Data Analytics. Lecture Notes in Networks and Systems, vol 281.
of our framework needs to implement real-time deployment Springer, Singapore. https://doi.org/10.1007/978-981-16-4244-9
capabilities as well as expand the usage of diverse datasets in [20] Datta, P., Rani, S., Koundal, D. (2020). ”Detection of Eye Ailments
Using Segmentation of Blood Vessels from Eye Fundus Image”. In:
combination with interpretability features that improve doctor Singh, P., Kar, A., Singh, Y., Kolekar, M., Tanwar, S. (eds) Proceedings
understanding of predictions. Current research efforts show of ICRIC 2019. Lecture Notes in Electrical Engineering, vol 597.
significant advancement in creating automated disease detect Springer, Cham. https://doi.org/10.1007/978-3-030-29407-63 7
[21] Mannepalli, P.K., Baghela, V.D.S., Agrawal, A., Johri, P., Dubey, S.S.,
systems that can expand health care availability worldwide. Parmar, K. (2024). ”Transfer learning for automated classification of eye
disease in fundus images from pretrained model”. Traitement du Signal,
Vol. 41, No. 5, pp. 2459-2470. https://doi.org/10.18280/ts.410520