Deep learning based binary classification of diabetic retinopathy images using transfer learning approach

Abstract

Objective

Diabetic retinopathy (DR) is a common problem of diabetes, and it is the cause of blindness worldwide. Detection of diabetic radiology disease in the early detection stage is crucial for preventing vision loss. In this work, a deep learning-based binary classification of DR images has been proposed to classify DR images into healthy and unhealthy. Transfer learning-based 20 pre-trained networks have been fine-tuned using a robust dataset of diabetic radiology images. The combined dataset has been collected from three robust databases of diabetic patients annotated by experienced ophthalmologists indicating healthy or non-healthy diabetic retina images.

Method

This work has improved robust models by pre-processing the DR images by applying a denoising algorithm, normalization, and data augmentation. In this work, three rubout datasets of diabetic retinopathy images have been selected, named DRD- EyePACS, IDRiD, and APTOS-2019, for the extensive experiments, and a combined diabetic retinopathy image dataset has been generated for the exhaustive experiments. The datasets have been divided into training, testing, and validation sets, and the models use classification accuracy, sensitivity, specificity, precision, F1-score, and ROC-AUC to assess the model's efficiency for evaluating network performance. The present work has selected 20 different pre-trained networks based on three categories: Series, DAG, and lightweight.

Results

This study uses pre-processed data augmentation and normalization of data to solve overfitting problems. From the exhaustive experiments, the three best pre-trained have been selected based on the best classification accuracy from each category. It is concluded that the trained model ResNet101 based on the DAG category effectively identifies diabetic retinopathy disease accurately from radiological images from all cases. It is noted that 97.33% accuracy has been achieved using ResNet101 in the category of DAG network.

Conclusion

Based on the experiment results, the proposed model ResNet101 helps healthcare professionals detect retina diseases early and provides practical solutions to diabetes patients. It also gives patients and experts a second opinion for early detection of diabetic retinopathy.

Introduction

The most common eye conditions that cause blindness or reduced vision are diabetic retinopathy, cataracts, glaucoma, and age-related macular degeneration (AMD) [1]. Diabetes is a condition that is common all over the world. Diabetes is ranked seventh among deadly diseases according to a World Health Organization (WHO) report [2–5]. It is noted that the number of cases and incidence of diabetes have been increasing over the past few decades, with an estimated 422 million people with diabetes disease [3–6]. It is noted that approximately 62 million Indians suffer from diabetic retinopathy with age 25–75, and it is expected that 102 million rise by 2030[1–5]. It is familiar to people with diabetes, and the retina is damaged due to high blood sugar [6]. The blood vessels leak, swell, and do not pass blood in the retina, causing abnormalities. Diabetic retinopathy damages the retina due to the complexities of diabetes mellitus and is the leading cause of blindness [7–10]. Diabetic Retinopathy has been classified into two stages named as (a) Proliferative Diabetic Retinopathy (PDR) and (b) Non-proliferative Diabetic Retinopathy (NPDR). Further, NPDR is divided into five classes: (a) Mild, (b) Moderate, and (c) severe [11–16]. The advanced form of diabetic retinopathy is called PDR, caused by the development of aberrant blood vessels or swelling in the retina [14]. The first stage of DR is known as mild NPR, during which the patient does not notice any changes in the vision or eye condition [15]. The blood vessel walls of the retina dilute and cause leakage, leading to microaneurysms, i.e., small lumps coming out from vessel walls [16–19]. Generally, a minute red spot or yellow circle that appears in the eye is called mild, moderate, or severe NDPR. Moderate NPDR is distinguished from mild NPDR by more significant damage to the retina's blood vessels. In moderate NPDR, the blood arteries develop microscopic balloons or swellings with microaneurysms, and fluid leaks from the retina or bleeds [20–24]. Severe NDPR is a serious problem characterized by a more damaged retina in blood vessels than mild and moderate NPDR [24–26].

The exhaustive literature survey shows that deep learning-based pre-trained networks with a transfer learning approach have been widely used for classifying diabetic retinopathy images [1–42]. Fundus retina images are widely used to detect abnormalities of diabetic retinopathy. In this work, a robust dataset of DR has been prepared to classify images into binary classes [27]. PDP and NPDP are considered unhealthy eye conditions, and non-diabetic Retinopathy is considered healthy eye fundus images of the eye [28]. In unhealthy eye images, retinal vessels become irregular in shape, size, and diameter, whereas healthy eye images have regular shape, size, and diameter [29–33]. Three benchmarked datasets of diabetic retinopathy images have been considered in this work for the experiment. A brief description of healthy eye and diabetic or unhealthy eye images of retinopathy is shown in Fig. 1.

Fig. 1.

A brief description of healthy eye and diabetic or unhealthy eye images of retinopathy

In the present work, pre-processing of DR images has been used to pre-process the raw DR images named as (a) resizing of DR images, (b) removing of Gaussian and salt and paper noise with a suitable filtering method [30–33]. The quality of DR images has been enhanced by selecting an optimal filtering algorithm in this work [33–37]. The optimal filtering algorithm has been chosen by pulling different filtering categories in the context of smoothening the homogeneous region and preserving the edges of the DR images [38]. This work uses structure and edge preservation index to evaluate the performance of the filtering algorithm using DR images [37–40]. Accordingly, in the present work, the pre-trained network has been divided into three categories named (a) series, (b) DAG (Residual and inception modules), and lightweight networks [35–42]. In this work, 20 pre-trained have been selected in the robust pull of three categories, i.e., Series, DAG & Lightweight, and fined-tuned the pre-trained networks using the transfer learning approach [40–42]. Accordingly, classification accuracy, sensitivity, specificity, precision, F1-score, and ROC-AUC are used to evaluate the network performance [37–42].

Table 1 shows a brief exhaustive literature review on the classification of diabetic retinopathy imaging using a deep learning-based pre-trained network and transfer learning approach.

Table 1.

A brief exhaustive literature review on the classification of diabetic retinopathy imaging using a deep learning-based pre-trained network and transfer learning approach

Investigator(s)	Year	Pre-processed Method	Dataset Name	No. of Images	DL Based Pre-Trained Model	Classifier
Gulshan et al. [1]	2016	Resizing and normalization	EyePACS & MESSIDOR-2	-	DCNN	Sen.-97.5%
Chandrakumar et al. [2]	2016	Resizing and Contrast Enhancement	EyePACS, Drive	-	DCNN	Acc.-94%
Zhou, L. et al. [3]	2017	Contrast Enhancement	Three Dataset	-	Self-Design Architecture	AUC-0.928
Dutta, S et al. [4]	2018	Contrast Enhancement	EyePACS (5 Class)	50000	VGG16	Acc.- 78.3%
Junjun, P. et al. [5]	2018	Contrast Enhancement	EyePACS (5 Class)	35, 126	ResNet18	Acc.- 78.4%
Kassani, S. H. et. al [6]	2019	Resizing and normalization	APTOS 2019 (5 Class)	3662	Xception	Acc.- 83.09%
Challa, U.K et al. [7]	2019	Gaussian filters	Kaggle dataset (Five Class)	33,000	Pre-Trained Networks	Acc. -86.64%
Qummar, S et al. [8]	2019	Scaling and Resizing	Kaggle dataset	5608	5 Pre-trained Network	Sp and F1 Score
Bhardwaj, C. et al. [9]	2020	Contrast Enhancement	MESSIDOR (4 Class)	1200	QIV Model (Inception-V3)	Acc.- 93.33%
Saxena, G. et al. [10]	2020	Resizing and Augmentation	EyePACS	88,702	InceptionResNet, ResNet and Inception	AUC-0.927
Yusaku Katada et al. [11]	2020	DA	EyePACS	35,126 (3508 Selected)	Inception v3	Sensitivity of 81.5% and 90.8%
Ali Usman et al. [12]	2020	Resizing, Data Augmentation	7 Online Dataset	2680	Inception v3, ResNet50 and Alex Net	Acc.- 85.2%
Wejdan L. Alyoubi et al. [13]	2021	CLAHE, Cropping, and DA	DDR and APTOS-2019	47,870	EfficientNetB0	Acc.-89%
Bhardwaj, C. et al. [14]	2021	Contrast Enhancement	MESSIDOR (4 Class)	1200	QEIRV‑2 Model (Inception-V3)	Acc.- 93.3%
Chen, P. N. et al. [15]	2021	Resizing and Grayscale	EyePACS	(88,702),	NASNet-Large	Acc.- 81.60% & 92.5%
San-Li Yi et al. [16]	2021	Resizing and Augmentation	APTOS 2019 (5 Class)	3662	RA-EfficientNet	Acc.- 93.55%
Z. Khan et al. [17]	2021	Scaling and Resizing	EyePACS	88,702	VGG16 and VGG-NiN	Sp.-91%
Sraddha Das et al. [18]	2021	Adaptive histogram equalization	DIARETDB1	-	CNN	Acc.- 98.7%
AbdelMaksoud et al. [19]	2022	Resizing and Augmentation	4 Dataset	39,301	E-DenseNet	Acc.- 91.3%
Kobat, S. G., et al. [20]	2022	Resizing	NDRD APTOS 2019	2355 and 3662	DenseNet201	Acc. -87.43% & 84.90%
Mungloo-Dilmohamud, Z et al. [21]	2022	Rescaling and Augmentation	APTOS 2019 (5 Class)	3662	VGG16, ResNet50 DenseNet169	Acc. -82%
Al-Omaisi Asia et al. [22]	2022	Cropping, Resizing, and Augmentation	XHO Dataset (5 Class)	1607	ResNet 50, 101 and VGG16	Acc.-80.88%
Sambit S. Mondal et al. [23]	2022	CLAHE and DA	APTOS 2019 (5 Class)	3662	ResNext and DenseNet	Acc.-86.08
Yasashvini, R. et al. [24]	2022	Weiner Filter	APTOS 2019	3662	ResNet & DenseNet	Acc.- 96.22%
Dayana, A. M. et al. [25]	2022	ADF	Local	-	AFU-NET	-
Oulhadj, M. et al. [26]	2022	Scaling and Resizing	APTOS 2019	3662	Densenet-121, Xception, Inception-v3 & Resnet-50	Acc. -85.28%
Jabbar M. K. et al. [27]	2022	CLAHE, DA & Resizing	EyePACS	35,126	VggNet	Acc.-96.6%
Menaouer, B. et al. [28]	2022	Scaling and Resizing	APTOS‑2019, Messidor-2 & Local public DR	5584	VGG 16 and 19	Acc.-90.6% and
Fayyaz, A. M. et al. [29]	2023	Resizing and Augmentation	ODIR (4 Class)	-	AlexNet and ResNet-101	SVM Acc.-93%
Dolly Das et al. [30]	2023	Resizing and Augmentation	EyePACS (5 Class)	35, 126	19 Pre-Trained	Acc.-79.11%
C. Mohanty et al. [31]	2023	Cropped and Resized	APTOS 2019 (5 Class)	3662	DenseNet 121	Acc.-97.30%
Pradeep Kumar Jena et al. [32]	2023	CLAHE	APTOS and MESSIDOR	3662 & 1200	Self-Design Architecture	Acc.—SVM (98.6% and 91.9%)
Bhimavarapu, U. et al. [33]	2023	CLAHE and Histogram Equalization	APTOS and Kaggle	3662 & 35,126	5 Pre-trained Network	Acc. – 98.32% & 98.71%
Islam, N. et al. [34]	2023	Gaussian Filter	APTOS, IDRiD	3662 & 516	Xception	APTOS—99.04% IDRiD -94.17%
Sajid, M. et al. [35]	2023	DA and Image Enhancement	Public Dataset	32,800	DR-NASNet	Acc.-96%
Alwakid, G. et al. [36]	2023	CLAHE & Data Augmentation	APTOS 2019	3662	DenseNet-121	Acc.-98.36%
Vijayan, M. et al. [37]	2023	Scaling	DDR, IDRiD, and APTOS	13,673, 516 & 3662	6 Pre-trained Network	Acc.-82.5%
Alwakid, G. et al. [38]	2023	CLAHE & DA	APTOS 2019	3662	DenseNet-121	-
Guefrachi, S et al. [39]	2024	Resizing and Augmentation	APTOS 2019	3662	Resnet152-V2	Acc.- 100%
Sunkari, S. et al. [40]	2024	Contrast and Brightness	APTOS and Kaggle	3662 & 35,126	3 Pre-trained Network	Acc.—93.51%
Macsik, P. et al. [41]	2024	CLAHE & DA	DDR and APTOS 2019	3662	Xception & EfficientNetB4	Acc
Shakibania Bu-Ali et al. [42]	2024	CLAHE & DA	APTOS 2019	3662	4 Pre-trained Network	Acc.-96.44

T & V—Training & Validation, ODIR- Ocular Intelligent Recognition collect, NDRD-New Diabetic Retinopathy Dataset, APTOS-Asia Pacific Tele-Ophthalmology Society, ADF- Anisotropic Diffusion Filter, AFU-NET-Attention-based Fusion Network, DA-Data Augmentation

From Table 1, it is observed that APTOS 2019, Diabetic Retinopathy Dataset, IDRiD, and EyePACS datasets were used in most of the studies with binary and multi-class classification using tuned pre-trained networks using transfer learning [35–42]. It is concluded that the highest accuracy has been achieved at 100% by Guefrachi S et al. [39]. It is noted that the interpretability of pre-trained networks improves the classification accuracy and explains the ability of DR detection [30–42]. It is also observed that only 3 out of 42 studies (7%) used pre-trained networks as feature extractors and classifiers. It is observed that SVM, KNN, and PNN classifiers have been used to classify features extracted by CNN architecture [1–42]. It is also noted that only 2 out of 42 studies (approx. 5%) used self-design CNN architecture to classify diabetic retinopathy images. It is also stated that only 6 out of 42 studies used binary class classification [35–42]. It is also noted that Gaussian filters, CLAHE, Data augmentation, histogram equalization, scaling, and resizing have been widely used as despeckling algorithms [37–42]. According to the literature review, noise is removed from an image by convolving it with a Gaussian kernel while maintaining the underlying structures [7]. It's an easy way to reduce Gaussian noise without sacrificing performance [7, 28, 36–42].

In the present work, the pre-trained networks have been divided into three categories: series, DAG (Directed Acyclic Graphs), and lightweight. Selecting a network in which category to classify medical images is crucial. This work provided a platform to classify DR images into binary classes by selecting the optimal pre-trained network with category. Accordingly, 20 pre-trained networks have been divided into three categories: series, DAG, and lightweight. These networks are used to classify DR images into binary. However, lightweight pre-trained networks provide effective, scalable, and readily available options for deploying deep learning models in resource-constrained situations. DAGs and series help manage and optimize complex workflows with apparent dependencies.

In the present work, the Optimal Denoising algorithm, Gaussian Filter, and resizing image size are done by optimizing performance evaluation parameters and maintaining the aspect ratio of the DR images. It is noted that the robust pull of the filtering and resizing algorithm has selected the denoising and resizing method. Structure and edge-preserving index (SEPI) [43–50] metrics have been evaluated for the performance of denoising filters. The Gaussian optimal denoising filter for DR images performed outstandingly in the DR image classification. In the present work, the effect of the denoising algorithm has been evaluated based on the performance parameter of the classification used in this work. The impact of the denoising algorithm has been reported in this work.

Workflow adopted for classification of diabetic retinopathy images

The workflow adopted in this work for classifying diabetic retinopathy images using fine-tuned pre-trained networks with a transfer learning approach is presented in Fig. 2.

Fig. 2.

The workflow adopted in this work for the classification of diabetic retinopathy images using fined tuned pre-trained networks with the transfer learning approach

The exhaustive literature review noted that very few studies based on binary class classification of DR images differentiate retina disease into two classes, as no sign of diabetic retinopathy and the sign of diabetic retinopathy has been reported by the researchers [1–42]. It is also concluded that the feature engineering step has been eliminated using deep learning or different types of pre-trained networks [44]. It is also noted that feature engineering shows significantly less result than deep learning. Convention feature extraction and selection process has been removed using a convolution-based network followed by ReLU, normalization, and max pooling [43–47].

From Fig. 2, it is noted that the classification of DR images involves different steps named (a) Benchmark diabetic Retinopathy Dataset Collection, (b) Data Pre-processing and Preparation Module, (c) Dataset Splitting for Training, Validation, and Testing Set, (d) Data Augmentation (e) Selection of pre-trained network, (f) Network training (g) Hyperparameters tuning using transfer learning, (h) evaluation parameters, (i) Validation and interpretation of the trained model and (j) Development and deployment.

Original benchmark diabetic retinopathy dataset collection

The dataset for diabetic retinopathy images has been collected from three benchmarked datasets consisting of annotated images by experienced ophthalmologists marked as healthy and unhealthy retinal images affected by diabetic disease [51–53]. The sample images of diabetic retinopathy in different stages are shown in Fig. 3.

Fig. 3.

Diabetic Retinopathy Stages (a) Normal retinal, (b) Mild Diabetic Retinopathy, (c) Moderate Diabetic Retinopathy, (d) Severe Diabetic Retinopathy, (e) Proliferative Diabetic Retinopathy, (f) Macular Edema

The first dataset available at Kaggle, the Diabetic Retinopathy Dataset (DRD—EyePACS) [51], consists of 2750 DR images, 1000 belonging to healthy classes and 1750 to unhealthy classes. The size of all images is 256 × 256.

The second dataset, IDRiD (Indian Diabetic Retinopathy Image Dataset) [52], consists of 513 DR images, 413 of which belong to healthy and 103 to unhealthy classes. The size of all images is 4288 × 2848.

The third dataset, APTOS (Asia Pacific Tele-Ophthalmology Society)-2019 [53], contains 3662 DR images collected from many participants in rural India. The dataset was prepared by India's Aravind Eye Hospital [35–42]. The fundus images have been taken over a long period in various settings and situations. The medical professionals examined and labeled the samples based on the standard provided by the International Clinical Diabetic Retinopathy Disease Severity Scale (ICDRDSS). It consists of 1805 samples with healthy and 1857 samples with unhealthy classes. The details of diabetic retinopathy datasets are shown in Table 2.

Table 2.

The Detail of Diabetic Retinopathy Datasets

Class	DRD-EyePACS Dataset [51]	IDRiD Dataset [52]	ATOS -2019 Dataset [53]	Combined Dataset
Healthy (Not Diabetic Retinopathy)	1000	168	1805	2973
Unhealthy (Diabetic Retinopathy)	1750	348	1857	3955
Total	2750	516	3662	6928

Table 2 notes that four experiments were performed in this work. Datasets DRD-EyePACS, IDRiD, and APOS-2019 have been considered for experiments 1, 2, and 3. Experiment 4 has been performed in this work, combining all three datasets and generating a robust dataset. The datasets used by medical professionals are not robust, so to solve this problem, all datasets have been combined into two classes, and a collective dataset has been prepared in this work. 6928 DR images contain 2973 images in the healthy class and 3955 DR images in the unhealthy class.

Data pre-processing and preparation module

In the present work, the data pre-processing and preparation module has been divided into three categories named as (a) Resizing of DR Images, (b) Image enhancement using suitable denoising algorithm, (c) Data augmentation, and (d) Dataset Splitting for Training, Validation and Testing Set.

Resizing of DR images

The exhaustive literature review concludes that the resizing of DR images is essential for the analysis and classification of diabetic retinopathy disease. It is also noted that deep learning-based pre-trained networks required the standard size of the image. To meet this requirement, the image size has been resized by 256 × 256 in this work. The DR image has been resized by maintaining the aspect ratio.

Image enhancement using suitable denoising algorithm

The present work has enhanced DR image quality using a pre-processing step. From the exhaustive literature review, denoising, contrast adjustment, and sharpening have been widely used for DR images to enhance the quality of images [1–42]. In the present work, an exhaustive poll of filtering and enhancement algorithms has been selected for the categories named (a) linear, (b) non-linear, (c) edge-preserving, and (d) contrast enhancement categories [7, 35–42]. It is observed that the Gaussian filter outperformed in all categories. This work used a Gaussian filter to pre-process DR images [7].

Dataset splitting for training, validation and testing set

Dataset splitting is an essential step in deep learning model development to evaluate the performance of the networks. Initially, the dataset was divided into training and testing subsets, and then the training set was split into validation data and training data. A brief description of diabetic retinopathy datasets splitting for training, validation, and testing set is shown in Table 3.

Table 3.

The brief description of diabetic retinopathy datasets splitting for training and testing

Dataset	DRD-EyePACS Dataset		IDRiD Dataset		ATOS -2019 Dataset		Combined Dataset
Class	Training	Testing	Training	Testing	Training	Testing	Training	Testing
Healthy (Not Diabetic Retinopathy)	800	200	134	50	1605	200	2505	450
Unhealthy (Diabetic Retinopathy)	1550	200	298	50	1657	200	3487	450
Total DR Image	2350	400	416	100	3262	400	5992	900
	2750		516		3662		6928

Table 3 reveals that the training data contains a large portion of the set, and the network learns patterns and relationships from this data. The deep learning hyperparameters are fine-tuned, and their training performance is evaluated using the validation set. It prevents overfitting the deep learning network and calculates the networks' performance during each epoch. The trained model's performance is calculated using the testing set. It acts as a neutral gauge of the model applied to untested data. The dataset is typically divided into three subsets in equal ratios depending on the dataset's size and the problem's specific requirements.

Data augmentation

Balancing a dataset for diabetic retinopathy using data augmentation involves creating additional samples of the minority class (e.g., severe diabetic retinopathy) to match the number of samples in the majority class (e.g., no diabetic retinopathy) [33–42]. Data augmentation techniques generate new samples by transforming existing images while preserving their semantic content [40]. The dataset has been randomly shuffled, and the DR image dataset ensures that each sample type is represented in each subset [41]. To overcome these problems, data augmentation balanced the training and validation data and achieved good network performance [30–40]. In the present work, data augmentation has been applied to the training set of all three datasets. The healthy and unhealthy class of DR images reached approximately 4000 of each class and each dataset. The combined dataset was prepared by mixing all classes, and a robust dataset was prepared to classify DR images. The number of multipliers as data augmentation operations with each dataset is shown in Table 4.

Table 4.

The number of multipliers as data augmentation operation with each dataset

Dataset	DRD-EyePACS Dataset		IDRiD Dataset		ATOS -2019 Dataset		Combined Dataset
Class	Training	Testing	Training	Testing	Training	Testing	Training	Testing
Healthy (Not Diabetic Retinopathy)	800 × 5 =  4000	200	118 × 34 = 4012	50	1605 × 2.5 = 4012	200	4000 + 4012 + 4012 = 12,024	450
Unhealthy (Diabetic Retinopathy)	1550*2.6 = 4030	200	298 × 13.5 = 4023	50	1657 × 2.40 = 3976	200	4030 + 4023 + 3976 = 12,029	450
Total	8030	400	8035	100	7988	400	24,053	900

Randomly shuffle the dataset to ensure each subset contains a representative sample of the entire dataset. This helps reduce the risk of any particular subgroup being biased. In classification tasks, preserving the class allocation between subsets is essential, mainly when working with imbalanced datasets. In the present work, class subsets have been balanced using data augmentation, ensuring that each subset has an equal distribution of classes of diabetic retinopathy images.

The number of augmentation transform operations used as a multiplier for different images is shown in Table 5.

Table 5.

The number of augmentation transform operations used as multipliers for different images

Datasets	Diabetic Retinopathy disease type	Rotation		Flipping		Flipping with Rotation				Translation	Multiplier	No. of Images	Total Images
		90o	180o	Horizontal	Vertical	90° H	180° H	90oV	180oV
DRD-EyePACS Dataset -1	Healthy (Not Diabetic Retinopathy)	✓	✓	✓	✓	✓	×	×	×	×	5	800	4000
IDRiD Dataset -2		✓	✓	✓	✓	✓	✓	✓	✓	-13 to 13	34	118	4012
ATOS -2019 Dataset -3		✓	✓	×	×	×	×	×	×	0.5 ×  (-1)	2.5	1605	4012
DRD-EyePACS Dataset -1	Unhealthy (Diabetic Retinopathy)	✓	✓	×	×	×	×	×	×	0.6 ×  (-1)	2.6	1550	4030
IDRiD Dataset -2		✓	✓	✓	✓	✓	✓	✓	✓	-3 to 3	13.5	298	4023
ATOS -2019 Dataset—3		✓	✓	×	×	×	×	×	×	0.4 ×  (-1)	2.4	1657	3976

V-Vertical, H-Horizontal

Splitting the diabetic retinopathy dataset into training, testing, and validation sets

The first step in developing a deep learning-based classification model is to divide the DR Dataset into training, testing, and validation sets to ensure the model performs efficiently when applied to new data. Table 6 provides a brief description of the dataset after data augmentation was applied to diabetic retinopathy images, with balanced training, validation, and testing sets, as well as the combined dataset.

Table 6.

The brief description of the dataset after data augmentation applied to diabetic retinopathy images with balanced training, validation, and testing sets

Table 6 shows that the diabetic retinopathy datasets are considered balanced if there are the same number of cases or samples in each class or category in the training, validation, and testing sets. In the present work, the combined dataset has been generated. In deep learning applications, particularly classification, the balanced dataset is an important step to prevent biasing the majority of classes and guarantee the representation of all classes during training. It is also noted that diabetic retinopathy dataset performance has been compared with classification accuracy, and the best combination has been selected based on the dataset and type of pre-trained networks. It is also revealed that the performance of each pre-trained network has been calculated based on the transfer learning approach, and the denoising effect has been reported.

Selection of pre-trained network and classification module

Selecting a pre-trained network for diabetic retinopathy depends upon the type of dataset, size of images, and the availability of pre-trained models. In the present work, a pre-trained network has been selected based on Series, DAG, and lightweight categories. Table 7 shows a brief description of each category. The selection of a pre-trained network is essential for classifying diabetic retinopathy images. This work divides the pre-trained network into Series, DAG, and Lightweight categories. The number of parameters is the criteria for dividing the pre-trained network into each category [44].

Table 7.

A brief category of selecting a pre-trained network based on each category

S.No	Name of the Pre-trained Networks	Type of Categories	Size of Images	Number of Parameters(Million)	Depth of the Network
1	AlexNet	Series	227 × 227 × 3	61.0	8
2	vgg16		224 × 224 × 3	138	16
3	vgg19		224 × 224 × 3	144	19
4	darknet19		256 × 256 × 3	20.8	19
5	darknet53		256 × 256 × 3	41.6	53
6	inceptionv3	DAG	299 × 299 × 3	23.9	48
7	densenet201		224 × 224 × 3	20.0	201
8	Resnet50		224 × 224 × 3	25.6	50
9	Resnet101		224 × 224 × 3	44.6	101
10	xception		299 × 299 × 3	22.9	71
11	inceptionresnetv2		299 × 299 × 3	55.9	164
12	nasnetlarge		331 × 331 × 3	88.9	-
13	SqueezeNet	Lightweight Networks	227 × 227 × 3	1.24	18
14	mobilenetv2		224 × 224 × 3	3.5	53
15	shufflenet		224 × 224 × 3	1.4	50
16	nasnetmobile		224 × 224 × 3	5.3	-
17	efficientnetb0		224 × 224 × 3	5.3	82
18	GoogleNet		224 × 224 × 3	7.0	22
19	googlenet-places365		224 × 224 × 3	7.0	22
20	resnet18		224 × 224 × 3	11.7	18

Table 7 shows that the pre-trained networks selected in Series, DAG, and lightweight categories are 5, 7, and 8, respectively. A total of 20 pre-trained networks have been selected, and the depth of the pre-trained network is important for classifying DR images.

Assessment parameters used in classification of DR images

From an exhaustive literature survey, it is noted that Accuracy, sensitivity, specificity, log loss, precision, F-score, overlapping error, boundary-based evaluation, etc., are the performance metrics used by different researchers to evaluate the detection algorithms of diabetic retinopathy. The present work uses Accuracy, Sensitivity, Specificity, Precision, and F1-Score metrics to calculate performance parameters by classifying DR images. The sample of the Confusion matrix from the classification of DR images is shown in Table 8.

Table 8.

The sample of the Confusion matrix from the classification of DR images

Healthy (Not Diabetic Retinopathy)	TN	FP
Unhealthy (Diabetic Retinopathy)	FN	TP
	Healthy (Not Diabetic Retinopathy)	Unhealthy (Diabetic Retinopathy)

Number of correctly detected DR images True positive (TP) and the number of DR images detected wrongly, false-negative (FN) instances that are incorrectly classified, and the numbers of positive FP (false positive) and FN (false negative).

Accuracy is the ratio of the total number of correctly identified to the total number of images taken for the classification. The formula used for the accuracy is shown in Eq. 1.

$$Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$ 1

Sensitivity or recall determines the measure of the correctness of patients with DR. The formula used for Sensitivity is shown in Eq. 2.

$$Sensitivity=\frac{\mathit{TP}}{TP+FN}$$ 2

It is calculated by dividing the number of patients diagnosed with diabetic retinopathy by the total number of affected patients.

Specificity is a ratio that determines whether a person is unaffected by the DR disease. The formulas for Precision and recall are shown in Eq. 4 and 5, and the formula for Specificity is shown in Eq. 3.

$$Specificity=\frac{\mathit{TN}}{FP+TN}$$ 3

Precision is the ratio of correctly classified DR images to the number of unhealthy DR images detected wrongly as the DR images.

$$Precision=\frac{\mathit{TP}}{FP+TP}$$ 4

The F-Score, also known as the F1-Score, is a metric for how accurate a model is on a given dataset. It is used to evaluate binary classification systems.

$$F-Score=2\times \frac{Precision\times Recall}{Precision+Recall}$$ 5

The harmonic mean of precision and recall calculates the regular F1 score.

Implementation details and number of experiments

The selected hyperparameters used to implement experiments are shown in Table 9.

Table 9.

The selected hyperparameters for implementation of experiments

Hyperparameters	Details
Learning Rate	10–4
Mini-batch Size	32
Maximum Epochs	30
Optimizer	Adam
The selected machine for implementation of experiments
Used Machine	Details
GPU	NVidia GEFORCE RTX 4060, 8 GB, 3072 CUDA CORE
Processor	12th Gen Intel Core i7 Processor
Operating System	Windows 11 Home

In the present work, four experiments have been drawn from each category of the pre-trained network. The experiments performed to classify DR images are shown in Table 10.

Table 10.

Experiments Performed in the classification of DR images

Experiment(s)	Details	Dataset
Experiment—1	Using Original DR images without augmentation	DRD-EyePACS, IDRiD, ATOS-2019, and Combined dataset
Experiment—2	Using Original DR images with augmentation	DRD-EyePACS, IDRiD, ATOS-2019, and Combined dataset
Experiment—3	Using pre-processed dataset images without augmentation	DRD-EyePACS, IDRiD, ATOS-2019, and Combined dataset
Experiment—4	Using pre-processed dataset images with augmentation	DRD-EyePACS, IDRiD, ATOS-2019, and Combined dataset

Performance evaluation of experiments

In the present network, 20 pre-trained networks have been used to calculate the performance of three datasets and combined three datasets for DR images based on healthy and unhealthy classes. The pre-trained network has been divided into series, DAG, and lightweight categories. Accuracy, Sensitivity, Specificity, precision, and F1-score have been calculated in this work. In the present work, original, pre-processed with augmentation, and without augmentation DR images have been used to calculate the performance of pre-trained network categories. The performance of each case is as follows:

Performance evaluation metrics using DRD-EyePACS dataset

Several factors such as choice of pre-trained network, data pre-processing, fine-tuning of the pre-trained network using optimal hyper-parameter, validation data, and evaluation metrics are taken into consideration while networks performance using series, DAG and Lightweight network architectures category using DRD-EyePACS Dataset.

Performance of series-based network architectures using DRD-EyePACS dataset

Table 11 shows the performance of series-based networks with and without augmentation using original and pre-processed DR images of the DRD-EyePACS Dataset.

Table 11.

The performance of series-based networks based on augmentation and without augmentation using original & pre-processed DR images of DRD-EyePACS Dataset

(a) Using Original DRD-EyePACS Dataset images
Network Name	Confusion Matrix		Without Augmentation					Confusion Matrix		After Augmentation
			ACC %	Sen	Sp	Pr	F1			ACC %	Sen	Sp	Pr	F1
AlexNet	136	64	65	0.62	0.68	0.66	0.64	176	24	85	0.82	0.88	0.87	0.85
	76	124						36	164
vgg16	140	60	67.5	0.65	0.70	0.68	0.67	180	20	86	0.82	0.90	0.89	0.85
	70	130						36	164
vgg19	152	48	73	0.70	0.76	0.74	0.72	186	14	89.5	0.86	0.93	0.92	0.89
	60	140						28	172
darknet19	130	70	66	0.67	0.65	0.66	0.66	182	18	86.5	0.82	0.91	0.90	0.86
	66	134						36	164
darknet53	144	56	71	0.70	0.72	0.71	0.71	176	24	87	0.86	0.88	0.88	0.87
	60	140						28	172
(b) Using Pre-processed DRD-EyePACS Dataset Images
Network Name	Confusion Matrix		ACC	Sen	Sp	Pr	F1	Confusion Matrix		ACC	Sen	Sp	Pr	F1
AlexNet	140	60	69	0.68	0.70	0.69	0.69	178	22	87	0.85	0.89	0.89	0.87
	64	136						30	170
vgg16	150	50	71	0.67	0.75	0.73	0.70	182	18	87.5	0.84	0.91	0.90	0.87
	66	134						32	168
vgg19	152	48	75	0.74	0.76	0.76	0.75	192	8	92.5	0.89	0.96	0.96	0.92
	52	148						22	178
darknet19	144	56	69	0.66	0.72	0.70	0.68	176	24	85	0.82	0.88	0.87	0.85
	68	132						36	164
darknet53	156	44	73	0.68	0.78	0.76	0.72	186	14	89.5	0.86	0.93	0.92	0.89
	64	136						28	172

Table 11 shows that the vgg19 network in the series category achieved the highest classification accuracy using the DRD-EyePACS Dataset. It is also observed that 92.5% accuracy has been achieved after the augmentation of the DRD Dataset. It is also noted that the vgg19-based network achieved 0.89, 0.96, 0.96, and 0.92 evaluation parameters named sensitivity, Specificity, Precision, and F1-Score, respectively. It is also noted that Individual Class Accuracy (ICA) for the healthy eye is 192 and ICA for an unhealthy eye is 178, which has been achieved using series-based network vgg19. The best result is marked in grey.

Performance of DAG-based network architectures using DRD-EyePACS dataset

Table 12 shows the performance of DAG-based networks with and without augmentation using original and pre-processed DR images.

Table 12.

The performance of DAG-based networks based on augmentation and without augmentation using original & pre-processed DR images of DRD-EyePACS Dataset

(a) Using Original DRD-EyePACS Dataset images
Network Name	Confusion Matrix		Without Augmentation					Confusion Matrix		After Augmentation
			ACC %	Sen	Sp	Pr	F1			ACC %	Sen	Sp	Pr	F1
inceptionv3	138	62	67.5	0.66	0.69	0.68	0.67	180	20	86.5	0.83	0.90	0.89	0.86
	68	132						34	166
densenet201	152	48	69.5	0.63	0.76	0.72	0.67	184	16	87.5	0.83	0.92	0.91	0.87
	74	126						34	166
Resnet50	152	48	72.5	0.69	0.76	0.74	0.72	182	18	88.5	0.86	0.91	0.91	0.88
	62	138						28	172
Resnet101	154	46	74.5	0.72	0.77	0.76	0.74	190	10	91.5	0.88	0.95	0.95	0.91
	56	144						24	176
xception	152	48	71	0.66	0.76	0.73	0.69	184	16	87	0.82	0.92	0.91	0.86
	68	132						36	164
inceptionresnetv2	150	50	69	0.63	0.75	0.72	0.67	180	20	85	0.80	0.90	0.89	0.84
	74	126						40	160
nasnetlarge	148	52	68.5	0.63	0.74	0.71	0.67	188	12	88.5	0.83	0.94	0.93	0.88
	74	126						34	166
(b) Using Pre-processed DRD-EyePACS Dataset Images
Network Name	Confusion Matrix		ACC	Sen		Pr	F1	Confusion Matrix		ACC	Sen		Pr	F1
inceptionv3	152	48	72.5	0.69	0.76	0.74	0.72	184	16	88.5	0.85	0.92	0.91	0.88
	62	138						30	170
densenet201	154	46	73.5	0.70	0.77	0.75	0.73	184	16	89	0.86	0.92	0.91	0.89
	60	140						28	172
Resnet50	158	42	75	0.71	0.79	0.77	0.74	188	12	91.5	0.89	0.94	0.94	0.91
	58	142						22	178
Resnet101	162	38	77.5	0.74	0.81	0.80	0.77	194	6	93.5	0.90	0.97	0.97	0.93
	52	148						20	180
xception	150	50	72	0.69	0.75	0.73	0.71	182	18	89.5	0.88	0.91	0.91	0.89
	62	138						24	176
inceptionresnetv2	148	52	70	0.66	0.74	0.72	0.69	184	16	89	0.86	0.92	0.91	0.89
	68	132						28	172
nasnetlarge	142	58	69.5	0.68	0.71	0.70	0.69	176	24	86.5	0.85	0.88	0.88	0.86
	64	136						30	170

From Table 12, it is concluded that the Resnet101 network in the category of DAG achieved the highest classification accuracy using the DRD-EyePACS Dataset. It is also observed that 93.5% accuracy has been achieved after the augmentation of the DRD Dataset. It is also noted that the Resnet101-based network achieved 0.90, 0.97, 0.97, and 0.93 evaluation parameters named sensitivity, Specificity, Precision, and F1-Score, respectively. It was also noted that individual class accuracy (ICA) for a healthy eye is 194 and ICA for an unhealthy eye is 180, which was achieved using the based network Resnet101. The best result is marked in grey.

Performance of lightweight based network architectures using DRD-EyePACS dataset

Table 13 shows the performance of Lightweight networks based on augmentation and without augmentation using original and pre-processed DR images of the DRD-EyePACS Dataset.

Table 13.

The performance of Lightweight networks based on augmentation and without augmentation using original & pre-processed DR images of the DRD-EyePACS Dataset

(a) Using Original DRD-EyePACS Dataset images
Network Name	Confusion Matrix		Without Augmentation					Confusion Matrix		After Augmentation
			ACC	Sen	Sp	Pr	F1			ACC	Sen	Sp	Pr	F1
SqueezeNet	142	58	69.5	0.68	0.71	0.70	0.69	178	22	85	0.81	0.89	0.88	0.84
	64	136						38	162
mobilenetv2	154	46	72	0.67	0.77	0.74	0.71	180	20	87.5	0.85	0.90	0.89	0.87
	66	134						30	170
shufflenet	158	42	75.5	0.72	0.79	0.77	0.75	188	12	91	0.88	0.94	0.94	0.91
	56	144						24	176
nasnetmobile	152	48	72	0.68	0.76	0.74	0.71	180	20	88.5	0.87	0.90	0.90	0.88
	64	136						26	174
efficientnetb0	140	60	67	0.64	0.70	0.68	0.66	176	24	85	0.82	0.88	0.87	0.85
	72	128						36	164
GoogleNet	138	62	64.5	0.60	0.69	0.66	0.63	170	30	83.5	0.82	0.85	0.85	0.83
	80	120						36	164
googlenet-places365	150	50	71	0.67	0.75	0.73	0.70	182	18	87	0.83	0.91	0.90	0.86
	66	134						34	166
resnet18	142	58	68	0.65	0.71	0.69	0.67	176	24	90.5	0.93	0.88	0.89	0.91
	70	130						14	186
(b) Using Pre-processed DRD-EyePACS Dataset Images
Network Name	Confusion Matrix		ACC %	Sen	Sp	Pr	F1	Confusion Matrix		ACC %	Sen	Sp	Pr	F1
SqueezeNet	150	50	71	0.67	0.75	0.73	0.70	180	20	87.5	0.85	0.90	0.89	0.87
	66	134						30	170
mobilenetv2	158	42	74	0.69	0.79	0.77	0.73	186	14	90.5	0.88	0.93	0.93	0.90
	62	138						24	176
shufflenet	160	40	76.5	0.73	0.80	0.78	0.76	192	8	94.5	0.93	0.96	0.96	0.94
	54	146						14	186
nasnetmobile	158	42	74	0.69	0.79	0.77	0.73	184	16	90.5	0.89	0.92	0.92	0.90
	62	138						22	178
efficientnetb0	150	50	71	0.67	0.75	0.73	0.70	178	22	86.5	0.84	0.89	0.88	0.86
	66	134						32	168
GoogleNet	142	58	68	0.65	0.71	0.69	0.67	180	20	86	0.82	0.90	0.89	0.85
	70	130						36	164
googlenet-places365	152	48	72.5	0.69	0.76	0.74	0.72	182	18	89.5	0.88	0.91	0.91	0.89
	62	138						24	176
resnet18	152	48	70	0.64	0.76	0.73	0.68	176	24	86.5	0.85	0.88	0.88	0.86
	72	128						30	170

From Table 13, it is concluded that the shufflenet network in the category of Lightweight achieved the highest classification accuracy using the DRD-EyePACS Dataset. It is also observed that 94.5% accuracy has been achieved after the augmentation of the DRD Dataset. It is also noted that the shufflenet-based network achieved 0.93, 0.96, 0.96, and 0.94 evaluation parameters named sensitivity, Specificity, Precision, and F1-Score, respectively. It was also noted that individual class accuracy (ICA) for the healthy eye is 192, and ICA for an unhealthy eye is 186, achieved using a series-based network shufflenet. The best result is marked in grey.

Performance evaluation metrics using IDRiD Dataset

Several factors such as choice of pre-trained network, data pre-processing, fine-tuning of the pre-trained network using optimal hyper-parameter, validation data, and evaluation metrics are taken into consideration while networks performance using series, DAG and Lightweight network architectures category using IDRiD Dataset.

Performance of series-based network architectures using IDRiD dataset

The performance of series-based networks with and without augmentation using original and pre-processed DR images of the IDRiD Dataset is shown in Table 14.

Table 14.

The performance of series-based networks based on augmentation and without augmentation using original & pre-processed DR images of IDRiD Dataset

(a) Using Original IDRiD Dataset images
Network Name	Confusion Matrix		Without Augmentation					Confusion Matrix		After Augmentation
			ACC %	Sen	Sp	Pr	F1			ACC %	Sen	Sp	Pr	F1
AlexNet	42	8	64	0.44	0.84	0.73	0.55	46	4	85	0.78	0.92	0.91	0.84
	28	22						11	39
vgg16	43	7	67	0.48	0.86	0.77	0.59	46	4	86	0.80	0.92	0.91	0.85
	26	24						10	40
vgg19	42	8	72	0.60	0.84	0.79	0.68	47	3	90	0.86	0.94	0.93	0.90
	20	30						7	43
darknet19	39	11	62	0.46	0.78	0.68	0.55	44	6	84	0.80	0.88	0.87	0.83
	27	23						10	40
darknet53	41	9	71	0.60	0.82	0.77	0.67	46	4	87	0.82	0.92	0.91	0.86
	20	30						9	41
(b) Using Pre-processed IDRiD Dataset Images
Network Name	Confusion Matrix		ACC %	Sen	Sp	Pr	F1	Confusion Matrix		ACC %	Sen	Sp	Pr	F1
AlexNet	45	5	68%	0.56	0.90	0.85	0.67	47	3	87	0.80	0.94	0.93	0.86
	22	28						10	40
vgg16	46	4	71%	0.50	0.92	0.86	0.63	47	3	88	0.82	0.94	0.93	0.87
	25	25						9	41
Vgg19	45	05	73%	0.56	0.90	0.85	0.67	48	2	92	0.88	0.96	0.96	0.92
	22	28						6	44
darknet19	38	12	66%	0.56	0.76	0.70	0.62	46	4	88	0.84	0.92	0.91	0.88
	22	28						8	42
darknet53	46	4	72%	0.52	0.92	0.87	0.65	45	5	89	0.88	0.90	0.90	0.89
	24	26						6	44

Table 14 shows that the vgg19 network in the series category achieved the highest classification accuracy using the IDRiD Dataset. It is also observed that 92% accuracy has been achieved after the augmentation of the IDRiD Dataset. It is also noted that the vgg19-based network achieved 0.88, 0.96, 0.96, and 0.92 evaluation parameters named sensitivity, Specificity, Precision, and F1-Score, respectively. It was also noted that individual class accuracy (ICA) for the healthy eye is 48 and ICA for an unhealthy eye is 44, which was achieved using series-based network vgg19. The best result is marked in grey.

Performance of DAG-based network architectures using IDRiD dataset

The performance of DAG-based networks with and without augmentation using original and pre-processed DR images of the IDRiD Dataset is shown in Table 15.

Table 15.

The performance of DAG-based networks based on augmentation and without augmentation using original & pre-processed DR images of IDRiD Dataset

(a) Using Original IDRiD Dataset images
Network Name	Confusion Matrix 100		Without Augmentation					Confusion Matrix		After Augmentation
			ACC	Sen	Sp	Pr	F1			ACC	Sen	Sp	Pr	F1
inceptionv3	40	10	67	0.54	0.80	0.73	0.62	46	4	88	0.84	0.92	0.91	0.88
	23	27						8	42
densenet201	42	8	69	0.50	0.84	0.76	0.60	45	5	87	0.84	0.90	0.89	0.87
	25	25						8	42
Resnet50	41	9	70	0.58	0.82	0.76	0.66	47	3	88	0.82	0.94	0.93	0.87
	21	29						9	41
Resnet101	44	6	72	0.56	0.88	0.82	0.67	47	3	90	0.86	0.94	0.93	0.90
	22	28						7	43
xception	43	7	69	0.52	0.86	0.79	0.63	42	8	85	0.86	0.84	0.84	0.85
	24	26						7	43
inceptionresnetv2	38	12	67	0.58	0.76	0.71	0.64	45	5	84	0.78	0.90	0.89	0.83
	21	29						11	39
nasnetlarge	37	13	68	0.62	0.74	0.70	0.66	46	4	86	0.80	0.92	0.91	0.85
	19	31						10	40
(b) Using Pre-processed IDRiD Dataset Images
Network Name	Confusion Matrix		ACC	Sen	Sp	Pr	F1	Confusion Matrix		ACC	Sen	Sp	Pr	F1
inceptionv3	41	9	68	0.54	0.82	0.75	0.63	46	4	89	0.86	0.92	0.91	0.89
	23	27						7	43
densenet201	43	7	71	0.56	0.86	0.80	0.66	47	3	90	0.86	0.94	0.93	0.90
	22	28						7	43
Resnet50	43	7	72	0.58	0.86	0.81	0.67	45	5	90	0.90	0.90	0.90	0.90
	21	29						5	45
Resnet101	44	6	73	0.58	0.88	0.83	0.68	47	3	92	0.90	0.94	0.94	0.92
	21	29						5	45
xception	44	6	70	0.52	0.88	0.81	0.63	47	3	88	0.82	0.94	0.93	0.87
	24	26						9	41
inceptionresnetv2	40	10	69	0.58	0.80	0.74	0.65	43	7	86	0.86	0.86	0.86	0.86
	21	29						7	43
nasnetlarge	39	11	70	0.62	0.78	0.74	0.67	47	3	88	0.82	0.94	0.93	0.87
	19	31						9	41

From Table 15, it is concluded that the Resnet101 network in the category of DAG achieved the highest classification accuracy using the IDRiD Dataset. It is also observed that 92% accuracy has been achieved after the augmentation of the IDRiD Dataset. It is also noted that the Resnet101-based network achieved 0.90, 0.94, 0.94, and 0.92 evaluation parameters named sensitivity, Specificity, Precision, and F1-Score, respectively. It was also noted that individual class accuracy (ICA) for a healthy eye is 47, and ICA for an unhealthy eye is 45, achieved using a series-based network, Resnet101. The best result is marked in grey.

Performance of Lightweight based network architectures using IDRiD dataset

Table 16 shows the performance of Lightweight networks based on augmentation and without augmentation using original and pre-processed DR images of the IDRiD Dataset.

Table 16.

The performance of Lightweight pre-trained networks based on augmentation and without augmentation using original & pre-processed DR images of IDRiD Dataset

(a) Using Original IDRiD Dataset images
Network Name	Confusion Matrix100		Without Augmentation					Confusion Matrix		After Augmentation
			ACC %	Sen	Sp	Pr	F1			ACC %	Sen	Sp	Pr	F1
SqueezeNet	41	9	68	0.54	0.82	0.75	0.63	44	6	83	0.78	0.88	0.87	0.82
	23	27						11	39
mobilenetv2	44	6	73	0.58	0.88	0.83	0.68	45	5	89	0.88	0.90	0.90	0.89
	21	29						6	44
shufflenet	44	6	74	0.60	0.88	0.83	0.70	46	4	90	0.88	0.92	0.92	0.90
	20	30						6	44
nasnetmobile	40	10	69	0.58	0.80	0.74	0.65	45	5	88	0.86	0.90	0.90	0.88
	21	29						7	43
efficientnetb0	38	12	65	0.54	0.76	0.69	0.61	41	9	82	0.82	0.82	0.82	0.82
	23	27						9	41
GoogleNet	39	11	65	0.52	0.78	0.70	0.60	43	7	83	0.80	0.86	0.85	0.82
	24	26						10	40
googlenet-places365	39	11	67	0.56	0.78	0.72	0.63	45	5	86	0.82	0.90	0.89	0.85
	22	28						9	41
resnet18	38	12	66	0.56	0.76	0.70	0.62	44	6	84	0.80	0.88	0.87	0.83
	22	28						10	40
(b) Using Pre-processed IDRiD Dataset Images
Network Name	Confusion Matrix		ACC %	Sen	Sp	Pr	F1	Confusion Matrix		ACC %	Sen	Sp	Pr	F1
SqueezeNet	40	10	70	0.60	0.80	0.75	0.67	45	5	86	0.82	0.90	0.89	0.85
	20	30						9	41
mobilenetv2	44	6	74	0.60	0.88	0.83	0.70	45	5	90	0.90	0.90	0.90	0.90
	20	30						5	45
shufflenet	45	5	75	0.60	0.90	0.86	0.71	46	4	91	0.90	0.92	0.92	0.91
	20	30						5	45
nasnetmobile	40	10	71	0.62	0.80	0.76	0.68	46	4	89	0.86	0.92	0.91	0.89
	19	31						7	43
efficientnetb0	39	11	66	0.54	0.78	0.71	0.61	44	6	84	0.80	0.88	0.87	0.83
	23	27						10	40
GoogleNet	40	10	68	0.56	0.80	0.74	0.64	43	7	85	0.84	0.86	0.86	0.85
	22	28						8	42
googlenet-places365	40	10	70	0.60	0.80	0.75	0.67	45	5	88	0.86	0.90	0.90	0.88
	20	30						7	43
resnet18	40	10	68	0.56	0.80	0.74	0.64	44	6	85	0.82	0.88	0.87	0.85
	22	28						9	41

From Table 16, it is concluded that the shufflenet network in the category of Lightweight achieved the highest classification accuracy using IDRiD Dataset. It is also observed that 91% accuracy has been achieved after the augmentation of the IDRiD Dataset. It is also noted that the shufflenet-based network achieved 0.90, 0.92, 0.92, and 0.91 evaluation parameters named sensitivity, Specificity, Precision, and F1-Score, respectively. It is also noted that Individual Class Accuracy (ICA) for the healthy eye is 46 and ICA for an unhealthy eye is 45, which has been achieved using series-based network shufflenet. The best result is marked in grey.

Performance evaluation metrics using the ATOS-2019 Dataset

Several factors such as choice of pre-trained network, data pre-processing, fine-tuning of the pre-trained network using optimal hyper-parameter, validation data, and evaluation metrics are taken into consideration while networks performance using series, DAG and Lightweight network architectures category using ATOS-2019 Dataset.

Performance of series-based network architectures using ATOS-2019 Dataset

The performance of series-based networks based with augmentation and without augmentation using original & pre-processed DR images of the ATOS-2019 Dataset is shown in Table 17.

Table 17.

The performance of series-based networks based on augmentation and without augmentation using original & pre-processed DR images of ATOS-2019 Dataset

(a) Using Original ATOS-2019 Dataset images
Network Name	Confusion Matrix		Without Augmentation					Confusion Matrix		After Augmentation
			ACC %	Sen	Sp	Pr	F1			ACC %	Sen	Sp	Pr	F1
AlexNet	140	60	66	0.62	0.70	0.67	0.65	172	28	85	0.84	0.86	0.86	0.85
	76	124						32	168
vgg16	141	59	68.5	0.67	0.71	0.69	0.68	180	20	87.5	0.85	0.90	0.89	0.87
	67	133						30	170
vgg19	152	48	72.5	0.69	0.76	0.74	0.72	188	12	91.5	0.89	0.94	0.94	0.91
	62	138						22	178
darknet19	142	58	68	0.65	0.71	0.69	0.67	180	20	87	0.84	0.90	0.89	0.87
	70	130						32	168
darknet53	141	59	67	0.64	0.71	0.68	0.66	177	23	86.5	0.85	0.89	0.88	0.86
	73	127						31	169
(b) Using Pre-processed ATOS-2019 Dataset Images
Network Name	Confusion Matrix		ACC %	Sen	Sp	Pr	F1	Confusion Matrix		ACC %	Sen	Sp	Pr	F1
AlexNet	145	55	67.5	0.63	0.73	0.69	0.66	179	21	86.5	0.84	0.90	0.89	0.86
	75	125						33	167
vgg16	145	55	70	0.68	0.73	0.71	0.69	184	16	88.5	0.85	0.92	0.91	0.88
	65	135						30	170
vgg19	153	47	73	0.70	0.77	0.75	0.72	191	9	92.5	0.90	0.96	0.95	0.92
	61	139						21	179
darknet19	145	55	69.5	0.67	0.73	0.71	0.69	182	18	88	0.85	0.91	0.90	0.88
	67	133						30	170
darknet53	143	57	68	0.65	0.72	0.69	0.67	180	20	87.5	0.85	0.90	0.89	0.87
	71	129						30	170

From Table 17, it is concluded that vgg19 networks in the category of series achieved the highest classification accuracy using the ATOS-2019 Dataset. It is also observed that 92.5% accuracy has been achieved after the augmentation of the ATOS-2019 Dataset. It is also noted that the vgg19-based network achieved 0.90, 0.96, 0.95, and 0.92 evaluation parameters named sensitivity, Specificity, Precision, and F1-Score, respectively. It also noted that individual class accuracy (ICA) for a healthy eye is 191 and ICA for an unhealthy eye is 179, which was achieved using series-based network vgg19. The best result is marked in grey.

Performance of DAG-based network architectures using ATOS-2019 dataset

Table 18 shows the performance of DAG-based networks with and without augmentation using original and pre-processed DR images of the ATOS-2019 Dataset.

Table 18.

The performance of DAG-based networks based on augmentation and without augmentation using original & pre-processed DR images of ATOS-2019 Dataset

(a) Using Original ATOS-2019 Dataset images
Network Name	Confusion Matrix		Without Augmentation					Confusion Matrix		After Augmentation
			ACC %	Sen	Sp	Pr	F1			ACC %	Sen	Sp	Pr	F1
inceptionv3	147	53	71.5	0.70	0.74	0.72	0.71	181	19	86	0.82	0.91	0.90	0.85
	61	139						37	163
densenet201	145	55	705	0.69	0.73	0.71	0.70	185	15	88	0.84	0.93	0.92	0.87
	63	137						33	167
Resnet50	146	54	71.5	0.70	0.73	0.72	0.71	185	15	90	0.88	0.93	0.92	0.90
	60	140						25	175
Resnet101	150	50	74	0.73	0.75	0.74	0.74	189	11	91.5	0.89	0.95	0.94	0.91
	54	146						23	177
xception	144	56	70	0.68	0.72	0.71	0.69	181	19	86.5	0.83	0.91	0.90	0.86
	64	136						35	165
inceptionresnetv2	140	60	67	0.64	0.70	0.68	0.66	170	30	83.5	0.82	0.85	0.85	0.83
	72	128						36	164
nasnetlarge	145	55	70.5	0.69	0.73	0.71	0.70	182	18	87.5	0.84	0.91	0.90	0.87
	63	137						32	168
(b) Using Pre-processed ATOS-2019 Dataset Images
Network Name	Confusion Matrix		ACC %	Sen	Sp	Pr	F1	Confusion Matrix		ACC %	Sen	Sp	Pr	F1
inceptionv3	146	54	72.5	0.72	0.73	0.73	0.72	185	15	87.5	0.83	0.93	0.92	0.87
	56	144						35	165
densenet201	146	54	71.5	0.70	0.73	0.72	0.71	187	13	89	0.85	0.94	0.93	0.88
	60	140						31	169
Resnet50	151	49	73	0.71	0.76	0.74	0.72	187	13	91	0.89	0.94	0.93	0.91
	59	141						23	177
Resnet101	152	48	74.5	0.73	0.76	0.75	0.74	193	7	94	0.92	0.97	0.96	0.94
	54	146						17	183
xception	146	54	71	0.69	0.73	0.72	0.70	185	15	88	0.84	0.93	0.92	0.87
	62	138						33	167
inceptionresnetv2	141	59	68	0.66	0.71	0.69	0.67	175	25	85	0.83	0.88	0.87	0.85
	69	131						35	165
nasnetlarge	147	53	72	0.71	0.74	0.73	0.72	187	13	89.5	0.86	0.94	0.93	0.89
	59	141						29	171

From Table 18, it is concluded that the Resnet101 network in the category of DAG achieved the highest classification accuracy using the ATOS-2019 Dataset. It is also observed that 94% accuracy has been achieved after augmentation of ATOS-2019 Dataset. It is also noted that the Resnet101-based network achieved 0.92, 0.97, 0.96, and 0.94 evaluation parameters named sensitivity, Specificity, Precision, and F1-Score, respectively. It is also stated that individual class accuracy (ICA) for the healthy eye is 193 and ICA for an unhealthy eye is 183, which was achieved using the based network Resnet101. The best result is marked in grey.

Performance of Lightweight based network architectures using ATOS-2019 dataset

Table 19 shows the performance of Lightweight networks based on augmentation and without augmentation using original and pre-processed DR images of the ATOS-2019 Dataset.

Table 19.

The performance of Lightweight networks based on augmentation and without augmentation using original & pre-processed DR images of ATOS-2019 Dataset

(a) Using Original ATOS-2019 Dataset images
Network Name	Confusion Matrix400		Without Augmentation					Confusion Matrix		After Augmentation
			ACC %	Sen	Sp	Pr	F1			ACC %	Sen	Sp	Pr	F1
SqueezeNet	143	57	70.5	0.70	0.72	0.71	0.70	180	20	86.5	0.83	0.90	0.89	0.86
	61	139						34	166
mobilenetv2	145	55	71	0.70	0.73	0.72	0.71	176	24	87	0.86	0.88	0.88	0.87
	61	139						28	172
shufflenet	149	51	72.5	0.71	0.75	0.73	0.72	187	13	91.5	0.90	0.94	0.93	0.91
	59	141						21	179
nasnetmobile	143	57	70	0.69	0.72	0.71	0.70	182	18	86.5	0.82	0.91	0.90	0.86
	63	137						36	164
efficientnetb0	135	65	66	0.65	0.68	0.66	0.65	170	30	83.5	0.82	0.85	0.85	0.83
	71	129						36	164
GoogleNet	130	70	64.5	0.64	0.65	0.65	0.64	165	35	81.75	0.81	0.83	0.82	0.82
	72	128						38	162
googlenet-places365	141	59	69	0.68	0.71	0.70	0.69	175	25	85.5	0.84	0.88	0.87	0.85
	65	135						33	167
resnet18	136	64	66	0.64	0.68	0.67	0.65	174	26	84	0.81	0.87	0.86	0.84
	72	128						38	162
(b) Using Pre-processed ATOS-2019 Dataset Images
Network Name	Confusion Matrix		ACC %	Sen	Sp	Pr	F1	Confusion Matrix		ACC %	Sen	Sp	Pr	F1
SqueezeNet	148	52	72	0.70	0.74	0.73	0.71	183	17	87.5	0.84	0.92	0.91	0.87
	60	140						33	167
mobilenetv2	149	51	72.5	0.71	0.75	0.73	0.72	180	20	89.5	0.89	0.90	0.90	0.89
	59	141						22	178
shufflenet	153	47	73.5	0.71	0.77	0.75	0.73	191	9	92.5	0.90	0.96	0.95	0.92
	59	141						21	179
nasnetmobile	145	55	71	0.70	0.73	0.72	0.71	185	15	89	0.86	0.93	0.92	0.89
	61	139						29	171
efficientnetb0	137	63	67	0.66	0.69	0.68	0.66	175	25	85.5	0.84	0.88	0.87	0.85
	69	131						33	167
GoogleNet	135	65	66	0.65	0.68	0.66	0.65	172	28	85	0.84	0.86	0.86	0.85
	71	129						32	168
googlenet-places365	143	57	70	0.69	0.72	0.71	0.70	180	20	87.5	0.85	0.90	0.89	0.87
	63	137						30	170
resnet18	139	61	67	0.65	0.70	0.68	0.66	175	25	85.5	0.84	0.88	0.87	0.85
	71	129						33	167

From Table 19, it is concluded that the shufflenet network in the category of Lightweight achieved the highest classification accuracy using the ATOS-2019 Dataset. It is also observed that 92.5% accuracy has been achieved after augmentation of ATOS-2019 Dataset. It is also noted that the shufflenet-based network achieved 0.90, 0.96, 0.95, and 0.92 evaluation parameters named sensitivity, Specificity, Precision, and F1-Score, respectively. It was also noted that individual class accuracy (ICA) for the healthy eye is 191 and ICA for an unhealthy eye is 179, which was achieved using series-based network shufflenet. The best result is marked in grey.

Performance evaluation metrics using the combined dataset

Several factors such as choice of pre-trained network, data pre-processing, fine-tuning of the pre-trained network using optimal hyper-parameter, validation data, and evaluation metrics are taken into consideration while network performance using series, DAG, and Lightweight network architectures category using Combined Dataset.

Performance of series-based network architectures using combined dataset

The performance of series-based networks-based augmentation and without augmentation using original & pre-processed DR images of the Combined Dataset is shown in Table 20.

Table 20.

The performance of series-based networks based on augmentation and without augmentation using original & pre-processed DR images of Combined Dataset

(a) Using Original Combined Dataset images
Network Name	Confusion Matrix		Without Augmentation					Confusion Matrix		After Augmentation
			ACC %	Sen	Sp	Pr	F1			ACC %	Sen	Sp	Pr	F1
AlexNet	318	132	65.33	0.60	0.71	0.67	0.63	415	35	91	0.90	0.92	0.92	0.91
	180	270						46	404
vgg16	324	126	67.88	0.64	0.72	0.69	0.67	409	41	90	0.89	0.91	0.91	0.90
	163	287						49	401
vgg19	346	104	72.66	0.68	0.77	0.75	0.71	428	22	94.66	0.94	0.95	0.95	0.95
	142	308						26	424
darknet19	311	139	66.44	0.64	0.69	0.67	0.66	421	29	92	0.90	0.94	0.93	0.92
	163	287						43	407
darknet53	326	124	69.22	0.66	0.72	0.71	0.68	410	40	90.1111	0.89	0.91	0.91	0.90
	153	297						49	401
(b) Using Pre-processed Combined Dataset Images
Network Name	Confusion Matrix		ACC %	Sen	Sp	Pr	F1	Confusion Matrix		ACC %	Sen	Sp	Pr	F1
AlexNet	330	120	68.77	0.64	0.73	0.71	0.67	419	31	92	0.91	0.93	0.93	0.92
	161	289						41	409
vgg16	341	109	70.55	0.65	0.76	0.73	0.69	413	37	91.11	0.90	0.92	0.92	0.91
	156	294						43	407
vgg19	350	100	73.88	0.70	0.78	0.76	0.73	435	15	96.22	0.96	0.97	0.97	0.96
	135	315						19	431
darknet19	327	123	68.88	0.65	0.73	0.70	0.68	425	25	93	0.92	0.94	0.94	0.93
	157	293						38	412
darknet53	345	105	70.66	0.65	0.77	0.73	0.69	421	29	91	0.88	0.94	0.93	0.91
	159	291						52	398

Table 20 shows that the vgg19 network in the series category achieved the highest classification accuracy using a Combined Dataset. It is also observed that 96.22% accuracy has been achieved after augmentation of Combined Dataset. It is also noted that the vgg19-based network achieved 0.96, 0.97, 0.97, and 0.96 evaluation parameters named sensitivity, Specificity, Precision, and F1-Score, respectively. It was also noted that individual class accuracy (ICA) for the healthy eye is 435 and ICA for an unhealthy eye is 431, which was achieved using series-based network vgg19. The best result is marked in grey.

Performance of DAG-based network architectures using combined dataset

Table 21 shows the performance of DAG-based networks based augmentation and without augmentation using original and pre-processed DR images of the Combined Dataset.

Table 21.

The performance of DAG-based networks based on augmentation and without augmentation using original & pre-processed DR images of Combined Dataset

(a) Using Original Combined Dataset images
Network Name	Confusion Matrix		Without Augmentation					Confusion Matrix		After Augmentation
			ACC %	Sen	Sp	Pr	F1			ACC %	Sen	Sp	Pr	F1
inceptionv3	325	125	69.2	0.66	0.72	0.70	0.68	415	35	89.55	0.87	0.92	0.92	0.89
	152	298						59	391
densenet201	339	111	69.6	0.64	0.75	0.72	0.68	414	36	89.11	0.86	0.92	0.92	0.89
	162	288						62	388
Resnet50	339	111	71.7	0.68	0.75	0.73	0.71	414	36	90.22	0.88	0.92	0.92	0.90
	143	307						52	398
Resnet101	348	102	0.74	0.71	0.77	0.76	0.73	426	24	93.55	0.92	0.95	0.95	0.93
	132	318						34	416
xception	339	111	70.3	0.65	0.75	0.73	0.69	417	33	91.33	0.90	0.93	0.92	0.91
	156	294						45	405
inceptionresnetv2	328	122	67.8	0.63	0.73	0.70	0.66	406	44	88.33	0.86	0.90	0.90	0.88
	167	283						61	389
nasnetlarge	330	120	69.3	0.65	0.73	0.71	0.68	416	34	90.11	0.88	0.92	0.92	0.90
	156	294						55	395
Using Pre-processed Combined Dataset Images
Network Name	Confusion Matrix		ACC %	Sen	Sp	Pr	F1	Confusion Matrix		ACC %	Sen	Sp	Pr	F1
inceptionv3	339	111	72	0.69	0.75	0.74	0.71	415	35	90.88	0.89	0.93	0.93	0.88
	141	309						72	378
densenet201	343	107	72.3	0.68	0.76	0.74	0.71	418	32	91	0.89	0.93	0.93	0.89
	142	308						66	384
Resnet50	352	98	73.7	0.69	0.78	0.76	0.73	420	30	91.44	0.90	0.93	0.93	0.91
	138	312						50	400
Resnet101	358	92	75.6	0.72	0.80	0.78	0.75	440	10	97.33	0.97	0.98	0.98	0.97
	127	323						14	436
xception	340	110	71.3	0.67	0.76	0.73	0.70	414	36	92.44	0.91	0.94	0.93	0.88
	148	302						66	384
inceptionresnetv2	329	121	69	0.65	0.73	0.71	0.68	402	48	89.77	0.89	0.91	0.91	0.87
	158	292						70	380
nasnetlarge	328	122	70.6	0.68	0.73	0.72	0.70	410	40	91.44	0.89	0.94	0.94	0.88
	142	308						68	382

From Table 21, it is concluded that the Resnet101 network in the category of DAG achieved the highest classification accuracy using the Combined Dataset. It is also observed that 97.33% accuracy has been achieved after augmentation of Combined Dataset. It is also noted that the Resnet101-based network achieved 0.97, 0.98, 0.98, and 0.97 evaluation parameters named sensitivity, Specificity, Precision, and F1-Score, respectively. It is also stated that individual class accuracy (ICA) for the healthy eye is 440, and ICA for an unhealthy eye is 436, achieved using the based network Resnet101. The best result is marked in grey.

Performance of lightweight-based network architecture using combined dataset

Table 22 shows the performance of Lightweight networks based on augmentation and without augmentation using original and pre-processed DR images of the Combined Dataset.

Table 22.

The performance of Lightweight networks based on augmentation and without augmentation using original & pre-processed DR images of Combined Dataset

(a) Using Original Combined Dataset images
Network Name	Confusion Matrix		Without Augmentation					Confusion Matrix		After Augmentation
			ACC %	Sen	Sp	Pr	F1			ACC %	Sen	Sp	Pr	F1
SqueezeNet	326	124	69.7	0.67	0.72	0.71	0.69	402	48	85.44	0.82	0.89	0.88	0.85
	148	302						83	367
mobilenetv2	343	107	71.6	0.67	0.76	0.74	0.70	401	49	87.44	0.86	0.89	0.89	0.87
	148	302						64	386
shufflenet	351	99	74	0.70	0.78	0.76	0.73	429	21	93.33	0.91	0.95	0.95	0.93
	135	315						39	411
nasnetmobile	335	115	70.7	0.67	0.74	0.72	0.70	407	43	87.55	0.85	0.90	0.90	0.87
	148	302						69	381
efficientnetb0	313	137	66.3	0.63	0.70	0.67	0.65	387	63	0.84	0.82	0.86	0.85	0.84
	166	284						81	369
GoogleNet	307	143	64.5	0.61	0.68	0.66	0.63	378	72	82.66	0.81	0.84	0.84	0.82
	176	274						84	366
googlenet-places365	330	120	69.6	0.66	0.73	0.71	0.69	402	48	86.22	0.83	0.89	0.89	0.86
	153	297						76	374
resnet18	316	134	66.8	0.64	0.70	0.68	0.66	394	56	86.88	0.86	0.88	0.87	0.87
	164	286						62	388
(b) Using Pre-processed Combined Dataset Images
Network Name	Confusion Matrix		ACC %	Sen	Sp	Pr	F1	Confusion Matrix		ACC %	Sen	Sp	Pr	F1
SqueezeNet	338	112	71.3	0.68	0.75	0.73	0.70	408	42	87.33	0.84	0.91	0.90	0.87
	146	304						72	378
mobilenetv2	351	99	73.3	0.69	0.78	0.76	0.72	411	39	90	0.89	0.91	0.91	0.90
	141	309						51	399
shufflenet	358	92	75	0.70	0.80	0.78	0.74	439	11	96.66	0.96	0.98	0.98	0.97
	133	317						19	431
nasnetmobile	343	107	72.3	0.68	0.76	0.74	0.71	415	35	89.66	0.87	0.92	0.92	0.89
	142	308						58	392
efficientnetb0	326	124	68.6	0.65	0.72	0.70	0.67	397	53	85.77	0.83	0.88	0.88	0.85
	158	292						75	375
GoogleNet	317	133	67.1	0.64	0.70	0.68	0.66	395	55	85.44	0.83	0.88	0.87	0.85
	163	287						76	374
googlenet-places365	335	115	71.1	0.68	0.74	0.73	0.70	407	43	88.44	0.86	0.90	0.90	0.88
	145	305						61	389
resnet18	331	119	68.4	0.63	0.74	0.71	0.67	396	54	87.55	0.87	0.88	0.88	0.88
	165	285						58	392

From Table 22, it is concluded that the shufflenet network in the category of Lightweight achieved the highest classification accuracy using the Combined Dataset. It is also observed that 96.66% accuracy has been achieved after augmentation of Combined Dataset. It is also noted that the shufflenet-based network achieved 0.96, 0.98, 0.98, and 0.97 evaluation parameters named sensitivity, Specificity, Precision, and F1-Score, respectively. It is also noted that Individual Class Accuracy (ICA) for the healthy eye is 439 and ICA for an unhealthy eye is 431, which has been achieved using series-based network shufflenet. The best result is marked in grey.

Result and discussion

When creating a model on limited resources, CNN image analysis uses transfer learning (TL), which involves moving feature weights from training on large image datasets to training on smaller datasets. It is possible to minimize the number of images in the target domain drastically. Generally, the model is trained on a smaller target dataset for feature extraction or fine-tuning based on the target domain's size and similarity, using pre-trained weights from ImageNet, a sizable dataset of natural images. However, if there is a significant difference between the source and target domains, the model's performance might not be recognized. Several pre-trained models have been used to classify diabetic retinopathy images. They show promising and different results and require less computational time. To solve this problem, 20 different pre-trained models have been divided into three categories: series, DAG, and lightweight. Three benchmark datasets have been collected for this work, and a combined dataset has been prepared for two classes. The best results have been achieved with weights sequentially transferred from ImageNet to prepared datasets. Our research showed that the deep learning method based on the ResNet101 network effectively distinguishes between normal retinal and DR images. The promising result shows that the resnet101-based pre-trained network achieved the highest classification accuracy using combined dataset. The amount of dataset used for training, the choice of hyperparameters, the pre-trained network category, and the specific dataset impact the level of precision of deep learning models compared. The present work uses Accuracy, Sensitivity, Specificity, Precision, and F1-Score metrics to calculate performance parameters by classifying DR images. The best-performing network has been selected based on classification accuracy for each category. Table 23 shows the comparison analysis of different networks with different categories. Figure 4 shows the ROC-AUC curve using the combined dataset for three categories.

Table 23.

The comparison analysis of different networks with different categories

Category	Dataset	Network Name	Confusion Matrix		ACC %	Sen	Sp	Pr	F1
Series	Vgg19	EyePACS Dataset	192	8	92.5	0.89	0.96	0.96	0.92
			22	178
		IDRiD Dataset	48	2	92	0.88	0.96	0.96	0.92
			6	44
		ATOS-2019	191	9	92.5	0.90	0.96	0.95	0.92
			21	179
		Combined Dataset	435	15	96.22	0.96	0.97	0.97	0.96
			19	431
DAG	Resnet101	EyePACS Dataset	194	6	93.5	0.90	0.97	0.97	0.93
			20	180
		IDRiD Dataset	46	4	91	0.90	0.92	0.92	0.91
			5	45
		ATOS-2019	193	7	94	0.92	0.97	0.96	0.94
			17	183
		Combined Dataset	440	10	97.33	0.97	0.98	0.98	0.97
			14	436
Lightweight	shufflenet	EyePACS Dataset	192	8	94.5	0.93	0.96	0.96	0.94
			14	186
		IDRiD Dataset	46	4	91	0.90	0.92	0.92	0.91
			5	45
		ATOS-2019	191	9	92.5	0.90	0.96	0.95	0.92
			21	179
		Combined Dataset	439	11	96.66	0.96	0.98	0.98	0.97
			19	431

Fig. 4.

The ROC-AUC curve using the combined dataset for three categories

The Final Result of the pre-trained network with classification accuracy using four datasets is shown in Table 24.

Table 24.

The Final Result of the pre-trained network with classification accuracy using four datasets

S.No	Network Name	Classification Accuracy
		EyePACS Dataset	IDRiD Dataset	ATOS-2019	Combined Dataset
1	AlexNet	87	87	86.5	92
2	vgg16	87.5	88	88.5	91.11
3	vgg19	92.5	92	92.5	96.22
4	darknet19	85	88	88	93
5	darknet53	89.5	89	87.5	91
6	inceptionv3	88.5	89	87.5	90.88
7	densenet201	89	90	89	91
8	Resnet50	91.5	90	91	91.44
9	Resnet101	93.5	92	94	97.33
10	xception	89.5	88	88	92.44
11	inceptionresnetv2	89	86	85	89.77
12	nasnetlarge	86.5	88	89.5	91.44
13	SqueezeNet	87.5	86	87.5	87.33
14	mobilenetv2	90.5	90	89.5	90
15	shufflenet	94.5	91	92.5	96.66
16	nasnetmobile	90.5	89	89	89.66
17	efficientnetb0	86.5	84	85.5	85.77
18	GoogleNet	86	85	85	85.44
19	googlenet-places365	89.5	88	87.5	88.44
20	resnet18	86.5	85	85.5	87.55

Tables 23 & 24 and Fig. 4 concluded that the combined dataset achieved the highest accuracy in all three categories: series, DAG, and Lightweight. It is observed that Vgg19, ResNet101, and shufflenet pre-trained networks achieved the highest accuracy of 96.22%, 97.33%, and 96.66% in series, DAG, and Lightweight categories. It is also noted that ResNet101 achieved the highest category in all cases. It is concluded that the ResNet101 pre-trained network in the DAG category is optimal for diabetic retinopathy disease detection in the early stage. The best result is marked in grey.

Conclusion

The exhaustive experiments concluded that the ResNet101-based pre-trained network in the category of DAG achieved the highest accuracy using combined datasets, i.e., DRD-EyePACS, IDRiD, and ATOS-2019. The significant advancement in the ResNet101 network achieves a balance between computing efficiency, depth, and accuracy. It is also observed that high accuracy, robustness, and efficiency have been achieved using ResNet101 in the category of DAG and observed as a powerful method for diagnosing and classifying diabetic retinopathy. It enhances the diagnosis and treatment in the early stage for the patient and can be used as real-time clinical practice. It is also noted that the implementation and training of ResNet101 require much processing compared to other networks and require high-quality labeled data. In this work, data was balanced using the augmentation technique, and balanced data was generated for robust model training. It is noted that 97.33% accuracy has been achieved using ResNet101. The proposed method can be used for routine time clinical practice.

Funding

This research received no external funding.

Data availability

The corresponding author has been authorized to access the data, and it will be shared for reasonable reasons.

Declarations

Ethics approval and consent to participate

Not Applicable.

Consent for publication

NA.

Conflicts of interest

There is no conflict of interest among the authors.