High efficiency and robustness are essential in the diagnosis of medical images. However, properly segmenting disease boundaries is challenging due to their arbitrary shape, color, and location. Deep learning models with low generalization ability are limited in handling segmentation and prediction of the severity of DR. The existing U-Net models have low interpretability and confidence in decision-making.

Despite high computational complexity, these existing systems present limited accuracy in categorizing the severity levels of DR. Hence, to tackle this challenge, we propose a Hybrid Inception U-Net with low computational complexity and a high generalization ability for accurate prediction and categorization of DR severity. Inception blocks are designed to extract discriminating features from the input by using varied convolutional filter sizes with various scales.

The primary aim of this study is to introduce a novel CNN architecture for the detection of DR, modeled on the U-Net, that uses inception block concatenations to pass information between layers while outperforming the current extraction techniques. The salient contributions of the proposed work are as follows:

Design and development of Hybrid Inception U-Net (HIUNET) to increase the efficiency in segmenting DR under variations against color, shape, and location. The proposed HIUNET indicates where dilated convolution ends, and the proposed model’s encoding part aims to enlarge the receptive field.

The existing U-Net systems are limited in accuracy in categorizing the various severity levels of DR. Conventional hand-crafted Deep Learning models are still inept in classifying the severity of DR, based on images, despite their high computational complexity. Though deep ensemble models yield better accuracy with high-resolution images, low-resolution images compromise its accuracy curtailing its potential in generalized disease prediction.

To cover this challenge, we have proposed a Hybrid Inception U-Net with limited computational complexity and high generalization ability to yield better accuracy in categorizing diabetic severity prediction.

Subsequent sections present a structured and detailed explanation of the model. While Section 2 is a systematic survey of conventional practices and deep learning approaches implemented in DR diagnosis, Section 3 presents the proposed HIUNET model for DR. Section 4 stipulates the evaluation of the proposed system, and Section 5 concludes this report.

Traditional image classification uses three primary approaches to encode scale, rotate, and illumination variations: Histogram of Oriented Gradients (HoG) [21], Scale-Invariant Feature Transform (SIFT) [22], and Local Binary Pattern (LBP) [23]. Until feature extraction, Hessian-matrix-based candidate selection was used [24], and after that, the Support Vector Machine (SVM) classifier was in use. A previous study extracted hybrid features for DR classification [25]. In another study, a multilayer perception neural network obtains the components for categorization [26]. Eventually, an algorithm for MA detection was developed for image analysis [27]. It attained 83.62% sensitivity for three images with 39 injuries. The detection of red lesions method has also been proposed, wherein the lesions were labeled as candidates [28]. The structures of the candidates’ blood vessels were subtracted to reduce false positives. It increased the sensitivity by 94% and specificity by 87% when tested on 89 images. However, the disadvantage of this approach was a longer computation time, as it took three minutes per image. Pre-processing techniques, such as correlation, have also been applied to identify bleeding images [29], with a sensitivity of 85%. The disadvantage of this model was a high false positive rate, with four false positives per image.

The identification of MA was accomplished when the traversal segment enrapt on maximum pre-processing pixels [30]. The set of features used for the Bayesian classification includes statistical measures and changes in cross-section orientation. MA was detected through two approaches [31]. Morphological operations and the classifier of Naive Bayes were used for coarse and fine segmentation, respectively. Eighteen MA features were obtained for classification, giving an 85.68% sensitivity, 99.99% specificity, a precision of 83.34%, and an accuracy of 99.99%. Some studies have shown that MAs can be detected by finding blobs (regions of interest) using two approaches [32,33]. One method was for MA extraction, and the other was for ensemble-based MA detection. In terms of performance, this model was competitive compared to other individual detectors. Another study proposed Radon transform (RT) and a multi-overlapping windows algorithm To detect MAs [34]. Here, three different databases were used for performance evaluation. The model gained 94% sensitivity and 75% specificity for the Mashhad dataset. However, for the second local database, the model exhibited 100% specificity and 70% specificity. Another study used automated technology to detect DR based on fundus images [35]. A retrospective analysis was done from the eye check DR screening project using non-mydriatic images. Referable DR (RDR) was detected with 84% sensitivity and 64% specificity. For detecting DR lesions, a previous study proposed a deep convolutional network analysis (DCNN) [36]. Another study presented a DCNN model using traditional machine learning algorithms, with 94% area under the curve, 93% sensitivity, and 87% specificity, determined using the Messidor-2 and E-Ophtha datasets [37]. The authors proposed a study for identifying DR [38]. Their sensitivity and specificity for RDR were 90.5% and 91.6%, respectively, while for Sight Threatening Diabetic Retinopathy (STDR) were 100% and 91.1%. To effectively detect microaneurysms, the authors developed a hybrid text/image method [39]. This model had a precision of 99.7% and a recall of 87.8%, which compares favorably to state-of-the-art algorithms. The identification of early DR has also been discussed [40]. In this paper, the author further extracted features using Principal Component Analysis (PCA) and Radon transform and classified the components using a hierarchical system of classifiers. The proposed method analysis was done on diaretDB1, and sensitivity of 92.32% and 88.06%, respectively, were achieved.

As the CNN AlexNet became popular in many functional applications, the focus shifted to CNN architectures [8]. Put forward a five-stage deep CNN as proposed for DR screening [41]. For detecting DR in the early stages, a deep CNN architecture with PCA is developed [42]. A CNN architecture was suggested by the authors in which data augmentation is done to identify features such as MA, EX, and HA in over 5,000 fundus images with 95% sensitivity and 75% accuracy [43]. The authors proposed a ResNet architecture, and it was reviewed for the arrangement of fundus images into normal and abnormal images, achieving 85% accuracy and 86% sensitivity [44]. The authors presented a Siamese network to detect the automated grading of DR in which input images are separated into small patches and extracted from the local-level features [45]. The authors presented an ensemble-based system dedicated to DR and diabetic macular edema. It achieved an accuracy of 90.07% for the five-class DR task and 96.85% accuracy for three-class diabetic macular edema [46]. Houby presented a CNN model using transfer learning to detect and stage classification and got accuracies of 86.5, 80.5, 63.5, and 73.7 for 2-class, 3-class, 4-class, and 5-class classifications, respectively [47]. A DCNN methodology with applied data augmentation obtained an accuracy of 94.5% [48]. The authors presented five CNN-based designs of AlexNet, GoogleNet, InceptionV4, Inception ResNetV2, and ResNetXt-50 models to detect DR [49]. The authors put forward a deep network patch-based approach and achieved a sensitivity of 0.940-95% CI [50]. Das, Biswas, and Bandyopadhyay presented a small and large dataset in which machine learning and deep learning are used to separate and analyze the features [51]. The authors presented an innovative approach to detect and classify DR using hybrid residual U-Net [52,53] The authors obtained an accuracy of 80.8% in an ensemble concept using the Kaggle dataset. The author presented a CAD system to detect DR [54]. The authors presented different models, such as VGG16, VGG19, and CNN, to detect DR with the help of illumination and field of view and achieved 80% to 83% accuracy [55]. A study presented a diagnosis of DR from fundus images, followed by pre-processing of the fundus images, data augmentation, and finally, application to the deep convolutional neural network. The team achieved an accuracy of 88.72% [56]. The performance of a proposed model was reported to achieve an accuracy of 83.09% using modified Xception architecture [57]. In a study, the authors proposed identifying the DR using pre-processing, convolution, rectified linear, pooling, and fully connected layers [58]. Potential treatment methods using fuzzy hypersoft mappings for identifying disease have also been used [59]. Comparative study of existing models can be shown in Table 2.

Generally, the images captured from different devices follow specific techniques developed by different organizations to produce the specification of images. Pre-processing is the principal technique that makes different devices familiar with the features of DR images. The process of pre-processing is as follows:

This study proposed a novel hybrid Inception U-Net (HIUNET). This architecture consists of a contracting and expansive path. In regular inception, the original block can be modified as Modified Inception block 1 and Modified Inception block 2. These blocks were further split into four inception modules, namely, Inception Module 1, Inception Module 2, Inception Module 3, and Inception Module 4, as shown in Fig. 3. The original Inception block comprises 1 × 1 convolution, 3 × 3 convolution, 5 × 5 convolution, and 3 × 3 max pooling. During training to reduce the calculation time and dimensions before each branch, a 1 × 1 convolution was added. The results obtained from the four branches are fused by acquiring feature information at different scales. During the convolution, adopting 1 × 1, 3 × 3, and 5 × 5 layers, set the convolution stride to 1 and the padding pixel to 0, 1, and 2, respectively.

After processing, the features thus obtained with similar dimensions were fused to get the final feature set. In this study, two modified Inception blocks, which further comprised three Inception Modules, were used. The inception module 1 consists of convolution, batch normalization, and activation. The inception module 2 consists of convolution, batch normalization, activation, convolution, batch normalization, and activation. Inception module 3 consists of convolution, batch normalization, activation, convolution, convolution, batch normalization, and activation. The input images are processed with 3 × 3 different Inception modules to obtain the feature map with increased channels. The modified Inception block-2 gets good results compared to Inception block-1 because, in Inception block-2, a 3 × 3 max pooling layer is used to reduce feature information and redundant information. This process ensures the extraction of meaningful feature information. This max pooling layer reduces overfitting and computational cost. Thus, Inception block-2 gets good results compared to Inception block-1.

Fig. 4 shows the functioning of the actual Inception U-Net architecture. In pixel-based image segmentation, U-Net is the prominent CNN architecture in the classification task. In the proposed work, inception blocks were integrated into U-Net and customized convolution layers to realize high accuracy in severity prediction. Initially, the width of the network is enlarged by adding parallel filters of different sizes to the primary U-Net model. In the proposed model, only parallel layers were used to reduce the number of parameters and computational complexity. In this proposed model, as shown in Fig. 5, we designed the downsampling/contraction and the upsampling/expansion paths for pattern recognition and returned the output in the input size. The model was customized to reduce the input using switch normalization. This technique automatically adapts different reduction operations, such as batch norm, instance norm, and layer norm. Appropriate padding was used by all convolutions. An activation function ReLu was used for intermediate convolutions, including inception and sigmoid for the output. The input images are processed with 3 × 3 sized filters in Inception modules to obtain the feature map with increased channels. The modified Inception block-2 produces good results compared to Inception block-1 because in Inception block-2 (MIU-Net), a 3 × 3 max pooling layer is used to reduce redundant feature information and extract meaningful features. Further, max pooling aids in reducing overfitting and computational costs. Thus, Modified Inception block-2 gets good results when compared to Inception block-1.

This study’s APTOS and KAGGLE datasets comprised 3662 and 35126 samples separated into five classes. From APTOS, 733 samples were used for performance assessment, while 2929 samples were used for model creation in Table 3. For KAGGLE, 7026 samples were used for performance assessment, while 28100 samples were used for model creation.

Hyperparameter tuning is selecting a collection of parameters to learn an algorithm. The learning process is managed using its value. To generalize data patterns, a model needs constraints, weights, and learning rates. These measurements are called hyperparameters. The performance of the proposed model is improvised through hyper-parameter tuning. In the proposed work, parameters are tweaked through manual search and optimization achieved through the following list of hyper-parameters in Table 4, along with the size of the input image, different optimizers, different learning rates, batch size, and the number of epochs.

The APTOS and KAGGLE datasets in this study were pre-processed. Pre-trained models VGG-16 and VGG-19 were applied. With a learning rate of 0.0001 for both extension to the Adaptive Moment estimation (Adamax) and Stochastic Gradient Descent (SGD), the VGG-16 model has 134,281,029 trainable parameters, with 91% and 81% accuracy rates, respectively. With learning rates of 0.001, 0.0001, and 0.00001 for Adamax, Adaptive Moment estimation (Adam), Nesterov-Accelerated Adaptive Moment Estimation (NADAM), and SGD, respectively, VGG-19 has 139,590,725 trained parameters, with an accuracy rate of 91%. There are 2,712,897 parameters in the IUNET Model. With learning rates of 0.0001 and 0.00001 for Adam and Adamax, respectively, IUNET exhibits a 94% accuracy rate. In contrast, with learning rates of 0.001, 0.001, and 0.00001 for NADAM, RMSProp, and SGD, respectively, it exhibits an 84% accuracy rate. The MIUNET model has 2,545,637 trainable parameters. With Adamax and NADAM, the MIUNET obtained an accuracy rate of 93% at a learning rate of 0.001, while with SGD, it reached an accuracy rate of 84% at a learning rate of 0.00001. We proposed a novel HIUNET model with 2,143,370 trainable parameters. With learning rates of 0.0001 and 0.0001 for Adam and SGD, the proposed HIUNET model achieved accuracy rates of 95% and 86%, respectively. The HIUNET model outperformed the pre-trained models in terms of accuracy. The APTOS dataset was applied to different models with different learning rates.

The models’ results are shown in Fig. 6. After pre-processing, the models were tested against the KAGGLE dataset. With learning rates of 0.0001 and 0.001 for NADAM and SGD optimizers, the VGG-16 model achieved accuracy rates of 90% and 84%, respectively. Furthermore, with learning rates of 0.0001 for Adamax and 0.0001 for the Root Mean Squared Propagation (RMSprop) optimizer, the VGG-19 model achieved accuracy rates of 90% and 83%, respectively. The IUNET model obtained a 90% accuracy rate with learning rates of 0.001 and 0.0001 for RMSprop and an 84% accuracy rate with a learning rate of 0.001 for NADAM.

The MIUNET model achieved a 90% accuracy rate with learning rates of 0.00001, 0.001, and 0.00001 for Adam, NADAM, and RMSprop, respectively, and 87% accuracy with a learning rate of 0.001 for SDG. The results for all the models are shown in Fig. 7.

Our proposed HIUNET model achieved a 92% accuracy rate with a learning rate of 0.0001 for Adam but only a 90% accuracy rate with learning rates of 0.00001, 0.001, and 0.00001 for Adamax, NADAM, and SGD, respectively. Again, our proposed HIUNET model outperformed the pre-trained models in terms of accuracy. The KAGGLE dataset was applied to different models with different learning rates.

Our proposed HIUNET model had 2,143,370 trainable parameters. Compared to VGG-16 (134,281,029), VGG-19 (139,590,725), IUNET (2,712,897), and MIUNET (2,545,637), our model had significantly fewer trainable parameters (23,544,837). Even with fewer parameters, our suggested HIUNET model beat the state-of-the-art techniques, obtaining an accuracy rate of 95% on the ATPOS dataset and 92% on the KAGGLE dataset. The performance of CNN models is based on their learning rate. A higher learning rate accelerates the learning process and increases the loss function, while a lower learning rate causes the loss function to decrease steadily.

An optimum learning rate must be chosen to minimize the loss function in classification problems. Adamax, as an optimizer, with a learning rate of 0.00001, yields a higher performance when compared to the remaining hyperparameters. The result obtained from the proposed HIUNET model is even better than the existing models. The APTOS and KAGGLE datasets were used to train our model, which attained 95% and 92% accuracy rates, respectively, outperforming the current models described in Table 5.

The proposed study has failed to predict correct class labels for the low-contrast images and could not pan out poorly segmented pupil images. However, there is a score to improve the generalization capability of the proposed work.