Noname manuscript No.

(will be inserted by the editor)

EfficientNetB3 Architecture for Diabetic

Retinopathy Assessment using Fundus Images

Abdelali ElMOUFIDI · Hind AMMOUN

the date of receipt and acceptance should be inserted later

Abstract Classification of the stages of diabetic retinopathy (DR) is consid-

ered a key step in the assessment and management of diabetic retinopathy.
Due to the damage caused by high blood sugar to the retinal blood ves-
sels, differ-ent microscopic structures can be occupied in the retinal area, such
as micro-aneurysms, hard exudate and neovasculari-zation. The convolutional
neural network (CNN) based on deep learning has become a promising method
for the analy-sis of biomedical images. In this work, representative images of
diabetic retinopathy (DR) are divided into five catego-ries according to the
professional knowledge of ophthalmologists. This article focuses on the use of
convolutional neu-ral networks to classify background images of DR according
to disease severity and on the application of pooling, Softmax Activation to
achieve greater accuracy. The aptos2019-blindness-detection database makes
it possible to verify the performance of the proposed algorithm.
Keywords Deep learning · Convolutional neural networks · EfficientNet ·
Diabetic retinopathy.

1 Introduction

1.1 Diabetes

Diabetes is a complete metabolic disorder that can lead to various vascular

complications in the body. In addition, when the disease coexists with other
general illnesses (hypertension, obesity, high cholesterol, etc.), the risk of eye
complications will also increase. Diabetes damages the small blood vessels in

ElMOUFIDI Abdelali and Hind AMMOUN

Sultan Moulay Slimane University, Faculty of Science and Technology, Beni Mellal
Tel.: +123-45-678910
Fax: +123-45-678910
E-mail: [Link]@[Link]
2 Abdelali ElMOUFIDI, Hind AMMOUN

the retina, which is the back layer of the eye. This is called diabetic retinopa-
thy. The retina converts the light and images that enter the eye into neural
signals sent to the brain [1].Diabetic retinopathy Diabetic retinopathy is a
common complication of diabetes that affects retinal function as shown in fig-
ure1. This pathology occurs when the blood vessels in the retina degenerate.
These damaged blood vessels can dilate, causing fluid (plasma, lipid, and /
or blood) to leak or even become blocked, leaving part of the retina without
blood flow. All of these phenomena that occur due to diabetes can cause pro-
gressive damage to the structure of the eyeball, leading to severe vision loss
or even blindness [2].

Fig. 1 Normal retina and retina with diabetic retinopathy

1.2 Phases of diabetic retinopathy

The existence and degree of the anomaly determine the severity of the dis-
ease. Identifying such manifestations as micro aneurysms, bleeding, new blood
vessel formation, and venous beads is the primary diagnostic process. Micro
aneurysms are blood clots 100 to 120 µm in size, usually round in shape. Blood
escaping from a ruptured blood vessel is called a hemorrhage. The abnormal
growth of tiny blood vessels is called neovascularization. Venous beading refers
to the expansion of the center of the vein adjacent to the occluded arteri-
ole. Patients with diabetic retinopathy are divided into patients with diabetic
no proliferative retinopathy (NPDR) and diabetic proliferative retinopathy
(PDR). In addition, depending on the severity of the disease, NPDR patients
can be divided into mild, moderate NPDR patients. and severe. The stages of
severity of diabetic retinopathy are described in figure 2. Pictures of the vari-
ous stages of diabetic retinopathy [3]. The parts of this article are classified as
follows: Part 2 provides a review of the research literature. Section 3 specifies
Title Suppressed Due to Excessive Length 3

the materials and methods proposed. Section 4 reviews the results and discus-
sion of the proposed system and the performance evaluation. Finally, section
5 specifies the full conclusion of the document.

Table 1 Different stages of DR.

Stages of DR ophthalmoscope recordings Corresponding label

in the proposed algorithm
Normal Without any abnormalities Healthy retina
Mild NPDR Presence of micro aneurysms only DR stage 1
Moderate NPDR Microaneurysms are present but DR stage 1
in a smaller amount to severe NPDR
Severe NPDR Venous beading in two or more regions DR stage 2
Prominent intraretinal microvascular
abnormality in one or more regions
PDR . Vitreous/pre-retinal hemorrhage DR stage3
. Neovascularization

Fig. 2 Stages of DR starting from healthy fundus image[25]

2 Related works

[!h] The computer-assisted detection system can accurately detect the level
of diabetic retinopathy, which makes it very popular among researchers. Over
the past decade, many studies have focused on the de-velopment of computer-
assisted systems capable of automatically detecting diabetic retinopathy using
traditional machine learning algorithms. Quellec et al. [4] used the traditional
KNN algorithm and two categories of optimal filters to achieve an AUC of
0.927. In addition, Sinthanayothin et al. [5] have pro-posed an automated
system that uses the KNN algorithm to detect the morphological features of
diabet-ic retinopathy, and its sensitivity and specificity are 80.21% and 70.66%,
respectively. In addition, in the article [6], neural networks are used to classify
diabetic retinopathy into three categories. They divided mild, moderate, and
severe diabetic retinopathy into accuracy rates of 82.6%, 82.6%, and 88.3%,
respec-tively. Larson et al. [7] The automatic diagnosis of diabetic retinopathy
4 Abdelali ElMOUFIDI, Hind AMMOUN

is shown on the background photo of the visibility threshold. They achieved

90.1% detection accuracy for real cases and 81.3% accu-racy for false cases.
Agurto et al. [8] used a multiscale decomposition based on amplitude modula-
tion and frequency modulation to distinguish images of diabetic retinopathy
from images of normal retina. In [9], the author reports that the area under
ROC is 0.98 for the texture characteristics and an accuracy of 99.17% for the
two-class classification using the wavelet transform with SVM. Jelinek et al.
[10] have proposed an automated detection of diabetic retinopathy by com-
bining the work of Spencer and the Cree system. de Spencer [11] and the Cree
system [10], which achieved a specificity of 90%. and a sensi-tivity of 85%
Abràmo et al. [12] for the automatic detection of diabetic retinopathy they
created an Eye-Check algorithm. They detected anomaly lesions with an AUC
of 0.839. Dupas et al. [13] have devel-oped a computer-assisted detection sys-
tem with KNN classification. Acharya et al. [14] uses the SVM classifier based
on bispectrum invariant characteristics to classify the five categories. Achieve
sensitivity, specificity and precision of 82%, 86% and 85.9% respectively. They
also studied four characteristics and obtained a classification accuracy of 85%,
a specificity of 86% and a sensitivity of 82%. Roy Chow-dhury and others. [15]
proposed a two-step classification method. First, it eliminates false alarms.
Then use GMM, KNN, and SVM for classification tasks. They achieved a sen-
sitivity of 100%, an AUC of 0.904 and a specificity of 53.16%. Deep learning
algorithms have become popular in recent years. Kaggle [16] has initiated sev-
eral competitions focused on automatic scoring for the detection of diabetic
retinopathy. Pratt et al. [17] introduced a CNN-based method that has sur-
passed even human experts in classifying advanced diabetic retinopathy. Kori
et al. [18] Use a set of connected ResNet and Densely networks to detect dia-
betic retinopathy and advanced stages of macular enema Torrey et al. [19] to
detect lesions on retinal fundus images, they developed a more interpretable
CNN model. In a similar study [20], an unbalanced weight map methodology
is used by Yang et al. [21] in order to emphasize the detection of lesions with
an AUC of 0.95. In [22] the VGG-16 and Inception-4 networks were used for
efficient classification of diabetic retinopathy.

3 Materials and the proposed method

3.1 Environment

The scripts are written in Python 3 on the Jupyter notebook on colab. No

configuration is required and computing resources, including the GPU, are
accessible free of charge.

3.2 Dataset

The dataset (aptos2019-blindness-detection) contains 3662 high resolution color

images labeled as train-ing set and 1928 color images that are not labeled as
Title Suppressed Due to Excessive Length 5

training set. In each group, the images are divided into 5 groups according to
the severity of the existing DR. Label 0 represents the control group. Labels 1
to 4 represent mild, moderate, severe and proliferative DR, respectively. The
following is a visual sum-mary of the diagnostic distribution as shown Figure
3. The group size is obviously unbalanced, with more than 1,800 images rep-
resenting the control group (label = 0), and less than 300 images in the most
severe category (label = 4). While this imbalance is expected in real world
data, it poses a problem for many machine learning models. In addition to the
unbalanced category, the image sizes in the dataset are also different as shown
in Figure 4.

Fig. 3 Distribution of diagnoses in the dataset

Fig. 4 Random sample of images present in the training set

6 Abdelali ElMOUFIDI, Hind AMMOUN

3.3 Method proposed

The objective of this research is to classify the fundus images with great pre-
cision in the different stages of diabetic retinopathy. classification of patients
in the different stages of diabetic retinopathy in a rapid manner is necessary.
Thanks to this research and to the application of an EfficientnetB3 architec-
ture, we try to increase the accuracy of the classification in the study of images
of diabetic retinopathy.

3.3.1 Image preprocessing

3.3.2 Resize Image

When defining the architecture of the model, which will be explained in detail
in a later section, one of the requirements is to define a fixed input form. When
performing this task, it is important to keep in mind that there is a balance
between speed of computation and loss of information. For clarification, when
the size of an image is reduced, information (pixels) is removed. Less infor-
mation means faster training times; however, it can also mean reduced overall
accuracy. an image size of 300 x 300 has been select-ed.

3.3.3 Dataset augmentation

The common problem in machine learning is unbalanced group size. During the
training of our models, the goal is to improve the precision during the following
iterations (eras). Since the model learns by find-ing patterns to distinguish
groups from one another, under-represented groups will be seen less often and
therefore will not be learned as well as their over-represented counterparts. To
mitigate the consequences of over / under-representation, data augmentation
is used. By adjusting specific parameters, applying random changes to the
original training images. These random changes are applied to each epoch,
which means that the model will train on ”different” images at each iteration.

3.3.4 Convolutional neural network

Convolutional neural network is a deep-learning neural network. It is a type of

acyclic (feed-forward) artificial neural network, in which the communication
pattern between neurons is inspired by the visual cortex of animals between
neurons is in-spired by the visual cortex of animals. Neurons in this region of
the brain are arranged so that they correspond to overlapping regions when
tiling the visual field1. Their operation is inspired by biological processes2,
they consist of a multilayer stack of perceptrons, the pur-pose of which is
to preprocess3 small amounts of information [23]. there are different CNN
architectures like Lenet, Alexnet, Googlenet, ConvNet, ResNet, etc. In this
research, we used the EfficienNet Architec-ture which is introduced by Google
AI.
Title Suppressed Due to Excessive Length 7

3.3.5 Proposed EfficienNet architecture

The EfficientNetB3 Convolutional Network is a network architecture where

provides a new scaling method that uniformly scales all dimensions of network
depth, width and resolution as shown in figure 5. This architecture applies
the grid search strategy to find the relationship between the different basic
net-work scaling dimensions under a fixed resource constraint. The could find
the appropriate scaling coeffi-cients for each of the dimensions to be scaled by
applying this strategy. Using these coefficients, the basic network was scaled
to the desired size [24]. By comparing EfficientNets with other existing CNNs
on ImageNet. Typically, by reducing the parameter size and FLOPS by an
order of magnitude, the Effi-cientNet model can achieve greater accuracy and
efficiency than existing CNNs. For example, in high precision systems, our
EfficientNet-B7 achieves a peak accuracy of 84.4% in the top 1 / 97.1% in
ImageNet top 5, at the same time 8.4 times smaller than previous Gpipe,
6.1 times faster CPU inference speed. Compared to The widely used ResNet-
50, EfficientNet-B4 uses similar FLOPS, and 7 improves top 1 accuracy from
76.3% of ResNet-50 to 82.6% (+ 6.3%) as shown in figure 6. Table 1and Figure
7 illustrate the network architecture of our proposed method that can detect
diabetic retinopathy by severi-ty classification. The resolution of the network
input layer is 300 x 300 pixels. We extend the architecture of EfficientNet
by adding GlobalAveragePooling2D, Flatten and Dropout by 0.5 to reduce
overfitting and Soft-Max layer with five classes. The architecture is optimized
for 3662 sample images in 40 epochs, the learning rate is 0.00005 and the
Adam optimizer is used for faster network optimization. Table 2 pro-vides a
summary of our training hyper parameter settings

Fig. 5 Model Scaling. (a) is a baseline network example; (b)-(d) are conventional scaling
that only increases one di-mension of network width, depth, or resolution. (e) is our proposed
compound scaling method that uniformly scales all three dimensions with a fixed ratio [24]
8 Abdelali ElMOUFIDI, Hind AMMOUN

Fig. 6 Model Size vs. Accuracy Comparison. EfficientNet-B0 is the baseline network devel-
oped by AutoML MNAS, while Efficient-B1 to B7 are obtained by scaling up the baseline
network. In particular, our EfficientNet-B7 achieves new state-of-the-art 84.4% top-1 / 97.1%
top-5 accuracy, while being 8.4x smaller than the best exist-ing CNN [24]

Fig. 7 Flow diagram of the proposed method

4 EXPERIMENTAL RESULTS

In this article, we have proposed a classified model by implementing the deep

learning approach; exact-ly, we built a model based on the EfficientNetB3
architecture and an aptos2019-blindness-detection dataset to train and test it.
In figure7. below, the screenshot shows part of the experimental results ob-
tained and fig. show the details of the accuracy obtained and the loss error.
The accuracy of the pro-posed model is 98.26%.
Title Suppressed Due to Excessive Length 9

Table 2 Modified EfficientNetB3 architecture for classification of diabetic retinopathy im-

ages

Layer (type) Output shape Parameter

Efficientnetb3(Functional) (None,10,10,1536) 10783535
Global average pooling2d 5(Pooling) (None, 1536) 0
Flatten-5 (Flatten) (None, 1536) 0
Dropout-5 (Dropout) (None, 1536) 0
Dense-4 (Dense) (None, 5) 7685

Table 3 The summary of the settings of our training hyper parameters.

Hyper parameters Values

Loss Function MSE

Optimizer Adam
Batch Size 32
Epoch 40

Fig. 8 This screenshot represents a part of the training process of our proposed model

Fig. 9 Training and validation accuracy Training and validation loss

4.1 Performance Evaluation on Severity Grading

A validation set containing 365 photographic background images was used to

evaluate the model. The model is evaluated based on the weighted average of
the macro mean, precision, recall, and f1 scores to understand model perfor-
mance using the researched single label classification method. Macro average
and weighted average assessment of precision, recall and f1 score Five classes.
The macro means are 0.94, 0.95, 0.94 and the weighted averages of 0.98, 0.98
and 0.98 of the precision, recall and f1 scores are recorded, respectively. Our
model shows the best precision, recall and f1 scores of No DR, which are 0.99,
1.00, and 0.99, respectively. The assessment of the macro-mean and weighted
10 Abdelali ElMOUFIDI, Hind AMMOUN

mean of the preci-sion, recall and f1 score by the diabetic retinopathy classi-
fication system is presented in Table3.

Table 4 Diabetic Retinopathy Classification system evaluation of Macro Average and

Weighted Average for precision, recall, and f1-Score evaluated over 365 sample images of
validation data

Classes Precision Recall F1-support

No DR 0.99 1.00 0.99
Mild DR 1.00 0.95 0.97
Moderate DR 1.00 0.97 0.98
Severe DR 0.85 0.89 0.87
Proliferative DR 0.87 0.93 0.90
MacroAverage 0.94 0.98 0.95
Weighted Average 0.98 0.94 0.98

The Receiver Operating Characteristic (ROC) curve is a graph showing

the performance of the classifica-tion model under all classification thresholds.
The curve plots the rate of true positives versus the rate of false positives. the
figure 11 illustrates the detailed results of the ROC graph of the classification
model

Fig. 10 The ROC graph of the classification model

Title Suppressed Due to Excessive Length 11

4.2 Confusion Matrix

The confusion matrix describes the performance of the classification model on

the validation set by comparing the actual label with the predicted label. The
matrix confusion of our model is evaluated as a one-label classification method
for five categories, as shown in Table 3. Each element of the confusion matrix
shows the comparison of each image between the actual label and the label.
predicted. Configure verification. Our model shows the best DR-free results
by making correct predictions on 180 images out of 180 images. Although the
correct prediction images for mild DR, moderate DR, severe DR, and prolif-
erative DR are 35, 97, 17, and 27 images out of 37, 100, 19, and 29, respectively.
Table 4 illustrates the detailed results of the confusion matrix.

Table 5 Confusion Matrix: The diagonal elements which represent the number of points
for which predicted label matches true label, while non-diagonal elements are those that are
wrongly classified by the classifier.

True Label / No DR Mild DR Moderate DR Severe DR Proliferative DR

Predicted Label
No DR 180 0 0 0 0
Mild DR 2 35 0 0 0
Moderate DR 0 0 97 1 2
Severe DR 0 0 0 17 2
Proliferative DR 0 0 2 2 27

5 Conclusion

By early detection and treatment of diabetic retinopathy, severe vision loss in

diabetic patients can be prevented. Deep learning is one of the most advanced
technologies for solving classification problems and provides better accuracy.
The efficient convolutional neural network architecture used to detect and
classify fundus images will help ophthalmologists further eradicate vision loss
caused by diabetic reti-nopathy. In this article, we propose an EfficientNetB3
model for the early detection of the five severities of diabetic retinopathy. We
have made several modifications to the EfficientNetB3 pre-trained network and
used pre-processing to improve network performance. Our network was trained
on APTOS 2019 dataset, which outperformed other state-of-the-art networks
in early-stage detection. The accuracy of the proposed model is 98.26
12 Abdelali ElMOUFIDI, Hind AMMOUN

References

1. Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional

neural networks. Com-mun ACM 2017;60(May (6)):84–90.
2. Hajeb Mohammad Alipour S, Rabbani H, Akhlaghi MR. Diabetic retinopathy grading
by digital curvelet trans-form. Comput Math Methods Med 2012;2012:1–11.
3. Doshi D, Shenoy A, Sidhpura D, Gharpure P. Diabetic retinopathy detection using deep
convolutional neural net-works. 2016 international conference on computing, analytics and
security trends (CAST), December; 2016.
4. G. Quellec, S. R. Russell, and M. D. Abramoff, “Optimal Filter Framework for Auto-
mated, Instantaneous Detection of Lesions in Retinal Images,” IEEE Trans. Med. Imaging,
vol. 30, no.2, pp. 523–533, Feb. 2011, doi: 10.1109/TMI.2010. 2089383.
5. C. Sinthanayothin, V. Kongbunkiat, S. Phoojaruenchanachai, and A. Singalavanija, “Au-
tomated screening system for diabetic retinopathy,” in 3rd International Symposium on
Image and Signal Processing and Analysis, 2003. ISPA 2003. Proceedings of the, vol. 2,
pp. 915–920, doi:10.1109/ISPA.2003.1296409.
6. S. C. Lee, E. T. Lee, Y. Wang, R. Klein, R. M. Kingsley, and A. Warn, “Computer
classification of nonprolifera-tive diabetic retinopathy.,” Arch. Ophthalmol. (Chicago, Ill.
1960), vol. 123, no. 6, pp. 759–64, Jun. 2005, doi: 10.1001/archopht.123.6.759.
7. S. C. Lee, E. T. Lee, Y. Wang, R. Klein, R. M. Kingsley, and A. Warn, “Computer
classification of nonprolifera-tive diabetic retinopathy.,” Arch. Ophthalmol. (Chicago, Ill.
1960), vol. 123, no. 6, pp. 759–64, Jun. 2005, doi: 10.1001/archopht.123.6.759.
8. N. Larsen, J. Godt, M. Grunkin, H. Lund-Andersen, and M. Larsen, “Automated De-
tection of Diabetic Retinopa-thy in a Fundus Photographic Screening Population,” Inves-
[Link] Vis. Sci., vol. 44, no. 2, p. 767, Feb. 2003, doi: 10.1167/iovs.02-0417.
9. C. Agurto et al., “Multiscale AM-FM Methods for Diabetic Retinopathy Lesion De-
tection,” IEEE Trans. Med. Im-aging, vol. 29, no. 2, pp. 502–512, Feb. 2010, doi:
10.1109/TMI.2009.2037146.
10. N. K, A. UR, N. KP, K. S, and B. SV, “Decision support system for diabetic retinopathy
using discrete wavelet transform,” Proc. Inst. Mech. Eng. Part H J. Eng. Med., vol. 227,
no. 3, pp. 251–261, Mar. 2013, doi: 10.1177/0954411912470240.
11. H. J. Jelinek, M. J. Cree, D. Worsley, A. Luckie, and P. Nixon, “An automated microa-
neurysm detector as a tool for identification of diabetic retinopathy in rural optometric
practice,” [Link]. Optom., vol. 89, no. 5, pp. 299–305, Sep. 2006, doi: 10.1111/j.1444-
0938.2006.00071. x.
12. T. Spencer, J. A. Olson, K. C. McHardy, P. F. Sharp, and J. V Forrester, “An image-
processing strategy for the segmentation and quantification of microaneurysms in fluo-
rescein angiograms of the ocular fundus.,” Comput. Bi-omed. Res., vol. 29, no. 4, pp.
284–302, Aug. 1996.
13. M. D. Abràmoff et al., “Automated Early Detection of Diabetic Retinopathy,” Oph-
thalmology,vol . 117, no. 6, pp. 1147–1154, Jun. 2010, doi: 10.1016/[Link].2010.03.046.
14. B. Dupas et al., “Evaluation of automated fundus photograph analysis algorithms for
detecting microaneurysms, haemorrhages and exudates, and of a computer-assisted diag
nostic system for grad-ing diabetic retinopathy,” Diabetes Metab., vol. 36, no. 3, pp.
213–220, Jun. 2010, doi: 10.1016/[Link].2010.01.002.
15. R. Acharya U, C. K. Chua, E. Y. K. Ng, W. Yu, and C. Chee, “Application of higher
order spectra for the identifi-cation of diabetes retinopathy stages.,” J. Med. Syst., vol.
32, no. 6, pp. 481–8,Dec. 2008.
16. S. Roychowdhury, D. D. Koozekanani, and K. K. Parhi, “DREAM: Diabetic Retinopa-
thy Analysis Using Machine Learning,” IEEE J. Biomed. Heal. Informatics, vol. 18, no.
5, pp. 1717–1728, Sep. 2014, doi: 10.1109/JBHI.2013.2294635.
17. “Kaggle: Your Home for Data Science.” [Online]. Available:
[Link] 31-Mar-2019].
18. H. Pratt, F. Coenen, D. M. Broadbent, S. P. Harding, and Y. Zheng, “Convolutional
Neural Networks for Diabetic Retinopathy,” Procedia Comput. Sci., vol. 90, pp. 200–205,
Jan. 2016, doi: 10.1016/[Link].2016.07.014.
19. A. Kori, S. S. Chennamsetty, M. S. K. P., and V. Alex, “Ensemble of Convolutional
Neural Networks for Auto-matic Grading of Diabetic Retinopathy and Macular Edema,”
Sep. 2018.
Title Suppressed Due to Excessive Length 13

20. L. Torrey, L. Torrey, and J. Shavlik, “Transfer Learning.”

21. Z. Wang and J. Yang, “Diabetic Retinopathy Detection via Deep Convolutional Net-
works for Discriminative Lo-calization and Visual Explanation,” Mar. 2017.
22. Y. Yang, T. Li, W. Li, H. Wu, W. Fan, and W. Zhang, “Lesion detection and Grading of
Diabetic Retinopathy via Two-stages Deep Convolutional Neural Networks,” May 2017.
23. Schmidt-Erfurth, U., Sadeghipour, A., Gerendas, B. S., Waldstein, S. M., & Bogunović,
H. (2018). Artificial intel-ligence in retina. Progress in retinal and eye research.
24. HU, Qingyong, YANG, Bo, XIE, Linhai, et al. Randla-net: Efficient semantic segmenta-
tion of large-scale point clouds. In: Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition. 2020. p. 11108-11117.
25. DOSHI, Darshit, SHENOY, Aniket, SIDHPURA, Deep, et al. Diabetic retinopathy
detection using deep convolu-tional neural networks. In: 2016 International Conference on
Computing, Analytics and Security Trends (CAST). IEEE, 2016. p. 261-266.