Indonesian J our nal of Electrical Engineering and Computer Science V ol. 37, No. 3, March 2025, pp. 1830 ∼ 1844 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v37.i3.pp1830-1844 ❒ 1830 RNN Diabetic framew ork f or identifying diabetic ey e diseases Arwa Albelaihi 1,2 , Dina M. Ibrahim 1,3 1 Department of Information T echnology , Colle ge of Computer , Qassim Uni v ersity , Buraydah, Saudi Arabia 2 School of Computing, Uni v ersity of Leeds, Leeds, United Kingdom 3 Department of Computers and Control Engineering, F aculty of Engineering, T anta Uni v ersity , T anta, Egypt Article Inf o Article history: Recei v ed Jan 9, 2024 Re vised Oct 3, 2024 Accepted Oct 7, 2024 K eyw ords: Bi-GR U Deep learning Diabetic e ye disease GR U Multi-class classication Recurrent neural netw orks Retinal fundus images ABSTRA CT Man y areas of image identication and classication for medical imaging di- agnostics ha v e greatly beneted from deep learning (DL). Diabetic retinopa- th y (DR) will become the most common cause of blindness w orldwide, mak- ing diabetes a major threat to public healt h. This research proposes an auto- mated identication system using deep recurrent neural netw orks (RNNs) to identify and classify four cate gories of diabetic e ye diseases: DR, cataract, glaucoma, and diabetic macular edema (DME). W e use three dif ferent model architectures based on RNN and their types, we called our proposed system RNN Diabetic frame w ork. These models are combined with one of the com- monly used architectures that support suf cient accurac y and speed for the model which is residual netw ork (ResNet)152V2. The three model architectures are RNN+ResNet152V2, g ated recurrent unit (GR U)+ResNet152V2, and bidi- rectional GR U (Bi-GR U)+ResNet152V2. The proposed models were assessed as collected datasets: DIARETDB0, DIARETDB1, messidor , HEI-MED, ocu- lar , and retina. A full analysis and e v aluation of these three deep RNN architec- tures are presented. The e xperiments sho wed that the Bi-GR U+ResNet152V2 model w ork ed better than the other tw o proposed models. In addition, we com- pare these three proposed models with the pre vious studies and nd that the pro- posed Bi-GR U+ResNet152V2 m odel achie v es the highest result s with accurac y equal to 99.8%, 98.1% sensiti vity , 98.6% specicity , 99.8% precision, 99.8% F1 score, and 99.8% areas under the curv e (A UC). This is an open access article under the CC BY -SA license . Corresponding A uthor: Dina M. Ibrahim Department of Information T echnology , Colle ge of Computer , Qassim Uni v ersity Buraydah, Saudi Arabia Email: d.hussein@qu.edu.sa 1. INTR ODUCTION Deep learning (DL) is utilized in man y areas to nd ne w w ays to solv e important problems, and it does well on tests that sort things into groups. Applications of articial intelligence (AI) tools and techniques are appropriate in the medical eld. One of the 21 st century’ s most potently re v olutionary technologies is AI. Strong machine learning (ML) tools and methods lik e deep con v olutional netw orks, generati v e adv ersarial netw orks (GANs), con v olutional neural netw orks (CNNs), recurrent neural netw orks (RNNs), and articial neural netw orks (ANNs) were used to bring about this change. DL has outperformed AI in speech recognition, picture characterisation, and natural language creation. J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 1831 In se v eral areas of medical imaging diagnostics, including image recognition and classication, DL has pro v en to be an e f fecti v e and signicant te chnique. By using fundus images to assess and diagnose e ye disorders, DL can be used to detect and classify e ye diseases, including diabetic e ye disease. Under the general cate gory of diabetic e ye illness are diabetic retinopath y (DR), cataract, glaucoma, and diabetic macular edema (DME) [1], [2]. Diabetic e ye disease can cause serious vision loss or e v en cause blindness in patients between the ages of 20 and 74. W ithout early detection of diabetic e ye disease, vision loss cannot be pre v ented. If identied early , 90% of diabetics can pre v ent diabetic e ye injury . The goal of this research is to impro v e diabetic e ye disease (DL) detection models. T o be emplo yed, fundus images of diabetic e ye diseases must be g athered in DL models. After that, se v eral image preparation methods are applied to the images. Utilizing pre-processed images, features are automa tically e xtracted, and analysis rules are learned and e v aluated in the literature with a focus on using ML for respiratory disease de- tection and classication. Incl uding the most important articles released from 2018 till the end of 2021. These ndings help researchers plan their w ork and pro vi d e a useful contrib ution. Therefore, we disco v ered that Ibrahim et al. [3] used transfer learning (TL) with the V GG19-CNN, ResNet152V2, ResNet152V2+g ated re- current uni t (GR U), and ResNet152 V2+bidirectional GR U (Bi-GR U) architectures for multiclass classication on a combination of chest X-ray and computed tomograph y (CT) scan images. Their research yielded a wealth of adv ancements that encouraged us to use GR U and Bi-GR U for the multiclassication of diabetic ocular disorders, particularly since these models are no v el to this eld. Additionally , Minarno et al. [4] found that using some augmentation procedures along with Ef cientNetB7 models increases the cl assication accurac y for diabetic e ye disorders. P atients with diabetes are among the most common illness groups in the w orld today . V isual loss can result from diabetic fundus disorders, which are the primary cause of blindness. V isual function is currently af fected by fundus illnesses such as DR, glaucoma, cataracts, and others [5]–[10]. CNN’ s outstanding feature learning capabilities ha v e allo wed it to e xcel in the re alm of fundus imaging. Numerous deep-learning architec- tures that ha v e dem o ns trated e xceptional performance in identifying certain diabetic e ye conditions ha v e been documented in the literature. T o the best of our kno wledge, the cate gorization paradigm of the four diabetes- related disorders (DR, DME, glaucoma, and cataract) had not been signicantly altered before the start of our study . Ho we v er , by the end of 2021, Sarki et al. [11] had pro vided a frame w ork for those four disorders’ classication. The y only used one CNN model in their research, though, and the dataset the y used re v ealed class disparities. Their accurac y of 81.33% is still re g arded as lo w when compared to the latest studies on DL classications. Pre vious studies on cataract disease did not suf ciently e xamine cataract prediction and classica tion [12]. T o cate gorize the cataract disease indi vidually , there w as also independent research. DME disease iden- tication w as suggested as a research g ap in [13] because it is e xtremely li k ely to indicate that the retina is de v eloping DR. This information helps researchers better understand the causes of retina-based disorders. The second DL model focuses on training datasets with unequal distrib utions of cl asses across v arious diseases and limited data. The ndings of the training set may not be accurate enough to be deemed satisf actory if it is small. It is tak en into c on s ideration as a potential remedy to apply traditional data augmentation techniques as enhancing methods. This study in v estig ated the ef fects of combining RNNs with residual netw orks (ResNet) in identifying diabetic e ye diseases. While earlier studies ha v e e xplored the impact of only ResNet, the y ha v e not e xplicitly addressed its inuence on increasing the perf o r mance of identifying diabetic e ye diseases. The main contrib utions of this study are: − Presenting the RNN diabetic frame w ork, a multi-classication DL model designed to detect and classify the four main e ye disorders associated with diabetes: glaucoma, cataract, DME, and DR. − Illustrating the inuence of combining RNNs with ResNet in identifying diabetic e ye diseas es by using ResNet152V2, GR U, and Bi-GR U. − Analyzing these DL architectures are pro vided using four classes of public fundus datasets (DR, DME, glaucoma, and cataract). − Ev aluating the suggested w ork by comparing it s performance metrics to those of v arious models gi v en in prior studies for detecting and classifying diabetic e ye diseases. The paper’ s remaining outline: section 2 summarizes the literature re vie w , whereas section 3 details the study’ s materials and techniques. Section 4 presents the comparati v e analysis and discussion. In section 5, the paper concludes. RNN Diabetic fr ame work for identifying diabetic e ye diseases (Arwa Albelaihi) Evaluation Warning : The document was created with Spire.PDF for Python.
1832 ❒ ISSN: 2502-4752 2. LITERA TURE REVIEW In this section, we present a fe w related deep neural netw ork studies on diabetic e ye problems. W e found that careful e v aluation of the architecture and the data input is necessary for creating an ef fecti v e neural netw ork classier . Man y studies ha v e been conducted in the literature, as sho wn in [14]–[17], re g arding using DL for classifying diabetic e ye illness utilizing fundus images. More people with diabetic e ye disease are presented in [18]–[22] adopt the TL approach. Rather than using random generation, the TL initializes the parameters based on prior learning. The rst layers pick up on e xtracting essential elements lik e edges, te xtures, and intuiti v ely . The top layers, analogous to blood v essels and e xudates, are more particular to the task. The TL is useful in cases where there is not enough data to train a neural netw ork from the be ginning [23]–[29]. Three CNN models DenseNet, ResNet50, and V GG16 were compared by P an et al. [30] on four dif ferent cate gories of DR lesions: microaneurysms, laser scars, leakages, and non-per fusion re gions (NP). 4,067 fundus uorescein angiograph y images from 435 e yes, 218 left e yes and 217 right e yes were collected into databases by them. In terms of computation, DR lesion identication and the procedure were determi ned to be ef cient. According to e xperimental results, DenseNet is a useful model for automatically recognizing and classifying retinal lesions in multi-label-classied fundus uorescein angi og r aph y (FF A) images. Ne v ertheless, because microaneurysms are often misclassied due to the widespread presence of uorescein, the procedure is not v ery accurate in identifying them. Samanta et al. [31] introduced a TL-based CNN architecture for color fundus photograph y , which performs relati v ely well in identifying DR from hard e xudates, blood v essels, and te xture on a much smaller dataset. The dataset w as trained using their model on four classes (no DR, mild DR, moderate DR, and proliferati v e DR) and reached a Cohen’ s Kappa score of 0.8836 on the v alidation set and 0.9809 on the training set. Their model uses se v eral architectures , such as Inception v1, Inception v2, Inception v3, Xception, V GG16, ResNet-50, DenseNet, and Ale xNet. Zhang et al. [32] presented a DeepDR frame w ork for DR detection. DeepDR acti v ely detects the e xistence and se v erity of DR from fundus images by ensemble learning and TL, using se v eral transfer and ensemble learning methods. Additionally , the authors introduced a ne w dataset for DR images called macula- centered retinal fundus images (13,767 images of 1,872 patients). The proposed netw ork achie v ed 97.5% sensiti vity with a specicity of 97.7%. Ho we v er , their model needs to be assessed with a more comple x and lar ger dataset. Using a small dataset, Badah et al. [23] de v eloped a model to detect hard e xudates using support v ector machine (SVM), a multilayer perceptron netw ork. The y found that CNN is not the best classier for hard e xudate, b ut it does perform well. Moreo v er , good outcome s in challenging e xtraction can be obtained by combining ef cient classiers with image-processing methods. A recent study demonstrated the application of CNN based on the ResNet152 model in conjunction with other ML classiers such as SVM, K-nearest neighbors (KNNs), Nai v e Bayes, multi-l ayer perceptron, decision tree, and random forest to identify infections causing glaucoma in patients. Furthermore, Dipu et al. [10], The ODIR dataset w as classied using four TL-based models. The ac- curac y of the models w as 97.23% for V GG-16, 90.85% for ResNet-34, 94.32% for MobileNetV2, and 93.82% for Ef cientNet. CNN can train discriminati v e features with an attention module for free, thanks to the design proposed in [20]. The y found multiple common labels for a patient’ s left and right fundus image pairings, which allo wed them to address the class balancing problem in the highly unbalanced ODIR dataset. Bhima v arapu and Battineni [25], an e n ha n c ed acti v ation function w as suggested for using fundus images to diagnose DR. This function automatically lo wers loss and processing time in the dif ferent CNN models. T o cate gorize DR images, four datasets DRIVE, CHASE, DIARETDB0, and Kaggle were tak en into consideration. The cur rent acti v ation functions used were scaled e xponential linear unit (SELU), rectied linear unit (ReLU), Sigmoid, and e xponential linear unit (ELU). On the Kaggle dataset, ResNet-152 has the greatest accurac y (99.41%). Butt et al. [18] suggests utilizing TL to e xtract fund picture features using ResNet-18 and GoogleNet models to correct class imbalance in a public dataset. Similarly , [19], [26] enhances picture quality with a h ybrid DL classier , utilizing ResNet50-based architecture. T able 1 sho ws a summary of studies that used dDL methods, including ho w the y w ork ed and ho w well the y classied images of the fundus that sho w diabetic diseases using v arious models of ResNet. 3. MA TERIALS AND METHODS Our mission is to e v aluate multiple DL models for detecting these four diabetes-related ocular prob- lems. Our dataset will be compiled using data sourced from six separate open-source databases. Once images Indonesian J Elec Eng & Comp Sci, V ol. 37, No. 3, March 2025: 1830–1844 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 1833 ha v e been collected, the y under go se v eral preprocessing steps. Once the dataset has been completed, the pre- trained ResNet152V2 model is used to classify each disease using three dif ferent models. There are also RNN models of the GR U and the Bi-GR U that are link ed to ResNet152V2. Hence, we named the system we sug- gested RNN Diabetic frame w ork. Figure 1 illustrates a summary of our proposed RNN Diabetic frame w ork methodology . T able 1. Summary of classication-based methods using dif ferent models of ResNet and their corresponding measures in pre vious studies Study Datas et DL model Performance [30] Kaggle ResNet50-I Acc=81.40% Spc=87.9% Sen=59.0% A UC=85.2% ResNet50-II Acc=90.97% Spc=92.7% Sen=97.6% A UC=92.31% ResNet50-III Acc=95.85% Spc=98.8% Sen=70.4% A UC=96.04% ResNet50-IV Acc=91.15% Spc=95.4% Sen=69.7% A UC=93.51% [25] DIRA TEDB0 ResNet-50 Acc=93.54% Pre=99.43% Sen=95.27% F1=98.42% Spe=98.32% A UC=89% ResNet-152 Acc=96.64% Pre=99.53% Sen=97.96% F1=99.15% Spe=98.79% A UC=97% Kaggle ResNet-50 Acc=94.64% Pre=95.74% Sen=94.24% F1=97.72% Spe=96.86% A UC=93% ResNet-152 Acc=99.41% Pre=99.89% Sen=98.28% F1=99.93% Spe=99.94% A UC=98% DRIVE ResNet-50 Acc=92.44% Pre=94.83% Sen=93.72% F1=95.88% Spe=95.27% A UC=93% ResNet-152 Acc=97.84% Pre=99.68% Sen=98.45% F1=99.57% Spe=99.26% A UC=94% CHASE ResNet-50 Acc=93.83% Pre=95.73% Sen=93.22% F1=94.68% Spe=96.95% A UC=96% ResNet-152 Acc=99.05% Pre=99.94% Sen=98.45% F1=99.89% Spe=99.59% A UC=97% [10] ODIR ResNet34 Acc=90.85% Pre=93.7% Sen=93.17% F1=92.65% [20] ODIR InceptionResNet Acc=97.38% F1=94.58% Sen=96.08% A UC=96.08% Pre=94.28% 3.1. Stage 1: datasets f or the study 3.1.1. Data collection Numerous sources of fundus images were accessed and compiled for our e xperiments. Included in t his group are DME, DR, cataract, and glaucoma. Initially , we chose 219 i mage datasets from DIARETDB0 and DIARETDB1 [33] for DR. Second, for DME with 151 images and 169 images, respecti v ely , we used Messidor [34] and HEI-MED [35]. The ODIR dataset [36] is the third source of the data set and includes information on se v eral classications of e ye diseases. F or cataracts, we used 312 images, and for glaucoma, we used 178 images. The nal dataset used for both glaucoma and cataracts is a retina dataset consisting of 99 and 100 pictures, respecti v ely , according to [37]–[49]. 412, 320, 277, and 219 im ages for cataract, DME, glaucoma, and DR, respecti v ely , are the results of the datasets that were g athered. The datasets that were g athered yielded a total of 1,228 images, which were then randomly split into training and v ali d a tion sets. The collected datasets yielded 412, 320, 277, and 219 pictures for cataract, DME, glaucoma, and DR, respecti v ely as sho wn in T able 2. From the datasets, 1,228 photos were randomly separated into training and v alidation sets. RNN Diabetic fr ame work for identifying diabetic e ye diseases (Arwa Albelaihi) Evaluation Warning : The document was created with Spire.PDF for Python.
1834 ❒ ISSN: 2502-4752 Figure 1. The methodology of our proposed RNN Diabetic frame w ork T able 2. Dataset details for DR, cataract, glaucoma, and DME classes Disease Dataset images T otal DIARETDB0 100 DR DIARETDB1 119 219 Messidor 151 DME HEI-MED 169 320 ODIR 312 Cataract Retina 100 412 ODIR 178 Glaucoma Retina 99 277 3.1.2. Dataset pr e-pr ocessing Preparing a dataset to satisfy the demands of a DL model is commonly achie v ed through pre-processing. Distinct pre-processing stages in the input images of our model. The dataset is di vided into four classications: DR, cataract, glaucoma, and DME, as the initial step in the process. JPEG, PNG, GIF , BMP , and TIFF are found among the image formats that T ensorFlo w does not support. The les’ e xtensions are merely for indication; the contents of the images are not altered in an y w ay . All image e xtensions are replaced with the JPG format in this research. These images will then be used as input in the step of pre-processing data. After reading, the y were do wnsizing the images to 224 × 224 , with a batch size of 32. This needs to be supplied because the pipeline handles batches of images that are all the same size. The y were of fering genuine v alue when it came to data shuf e decisions. T o get transformations equal to 100 and shuf ing, the y added optional random seeds. After that, the images should be normal ized and turned into arrays so that, the y may be used as input for the model’ s ne xt step. W ith the label mode cate gorial, a oat 32 tensor of shape (batch size, num classes) encodes the class inde x one-hot. T o ensure our dataset for DL model training include dif ferent photos. Finally , we randomly di vided it into 70% and 30% sections for training and v alidation. Sample fundus pictures for DR, cataract, glaucoma, and DME are sho wn in Figure 2. 3.2. Stage 2: deep r ecurr ent neural netw orks Se v eral supervised DL algorithms are used to create classication models in this research. T o detect and eliminate the four diabetic e ye illnesses, the y will test their abilities. Three models are used: RNN+ ResNet152V2, GR U+ResNet152V2, and Bi-GR U+ResNet152V2. W e e xplain each of the three models in the follo wing subsections. 3.2.1. Stage 2 model 1: RNN+ResNet152V2 deep model The R esNet152V2 model w as selected for this research as the model supports suf cient accurac y and speed. The model has a RNN using the ResNet152V2 model, a reshape layer , 2D global a v erage pooling, a dropout layer , and a dense layer with a Softmax acti v ation function to classi fy the image. Figure 3 pro vides more details on the model’ s architecture. 3.2.2. Stage 2 model 2: ResNet152V2 and GR U deep model As sho wn in Figure 4, this model consists of GR U follo wed by ResNet152V2. The primary benet of using GR U is that it allo ws for the maintenance of details that are irrele v ant to the prediction for an e xtended Indonesian J Elec Eng & Comp Sci, V ol. 37, No. 3, March 2025: 1830–1844 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 1835 period without eliminating them. ResNet152V2, a reshape layer , a GR U layer after layer with 256 units, a collapsed layer , a layer that is dense with 128 neurons, a layer with dropouts, and a dense layer use the function of Softmax acti v ation to classify the image into one of our four disease classications. Figure 2. Fundus image samples from the original dataset Figure 3. Stage 2 model 1: RNN+ResNet152V2 model architecture 3.2.3. Stage 2 model 3: ResNet152V2 and Bi-GR U deep model The sequential model of bi-GR U, follo wed by ResNet152V2, is the last proposed DL model in our proposed RNN Diabetic frame w ork. The model architecture is described in detail in Figure 5. The model consists of a ResNet152V2, follo wed by a reshape layer , a bi-GR U layer with 512 units, a dense layer , a dropout layer , and another dense layer , and it utilizes a Softmax acti v ation function to cat e gori ze the image as belonging to one of our four cate gories of disorders. RNN Diabetic fr ame work for identifying diabetic e ye diseases (Arwa Albelaihi) Evaluation Warning : The document was created with Spire.PDF for Python.
1836 ❒ ISSN: 2502-4752 Figure 4. Stage 2 model 2: GR U+ResNet152V2 model architecture Figure 5. Stage 2 model 3: Bi-GR U+ResNet152V2 model architecture 4. RESUL TS AND DISCUSSIONS 4.1. T raining parameters The models in this study were trained and assessed independently b ut with identical parameters and metrics. W e needed to increase the o v erall amount of data and epochs to adequately train DL models. T o achie v e the best results, we trained the m odels on v arious epochs and batch size settings. Using an optimizer and suitable t functions, our models were trained and v eried. Ev ery model ran through about 300 epochs of training, with a batch size of 32. The presented ndings are the highest v alidation v alues that could be ob- tained. The results were obtained by including the performance metric equations in the outputs of the resulting v alidation data. F or our suggested models, the (Adam) [7] optimizer w as used to obtain the optimal outcomes. T able 3 displays the learning rate (LR) v alues and optimizers used for all models. Indonesian J Elec Eng & Comp Sci, V ol. 37, No. 3, March 2025: 1830–1844 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 1837 T able 3. The training parameters of the models: learning rate v alue and optimizer Model Optimizer Learning rate RNN+ResNet152V2 Adam 0.00001 GR U+ResNet152V2 0.00001 Bi-GR U+ResNet152V2 0.00001 4.2. Early stopping (callback) Neural netw ork training requires epoch selection. T oo man y or too fe w epochs might o v ert or undert the training dataset. Set epochs with early stopping, model checkpoint, and other approaches. Our early stopping method in v olv es setting man y epochs and halting training when the impro v ement is not enough and matches our e xpectations. Early stopping sets an arbitrarily high number of training epochs and stops training when model performance on a hold-out v alidation dataset st ops impro ving. The early stopping callback from K eras lets you terminate training early . Automatic callbacks run code and interact with the training model. The v alidation dataset loss function for model optimization is calculated after each epoch. First e vidence of no impro v ement is not al w ays ideal to stop training. This is because the model may stall or w orsen before impro ving rapidly . T o postpone epochs without impro v ement, we set ”patience” to 60 in this e xperiment. 4.3. P erf ormance metrics The RNN Diabetic frame w ork w as assessed for diabetic e ye disease cate gorization utilizing measures such as accurac y , sensiti vity , precision, specicity , F1-score, and areas under the curv e (A UC). A confusion matrix is then added to each model. In (1), accurac y is the number of occurrences that can be predicted accurately gi v en the total number . Accur acy ( Acc ) = T p + T n T p + T n + F p + F n (1) where TP and TN are true positi v e and ne g ati v e parameters, whereas FP and FN are f alse positi v e and ne g ati v e v alues. In (2) gi v es recall as the number of samples that were both actually positi v e and projected to be positi v e. In (3) sho ws the precision, or positi v e predicti v e v alue, of all e xpected positi v e samples. In (4) gi v es the real ne g ati v e rate or specicity , which is the number of truly ne g ati v e samples anticipated to be ne g ati v e. Last, (6) sho ws the F1-score, or harmonic mean of precision and sensiti vity . S ensitiv ity ( S en ) = T p T p + F n (2) P r ecision ( P r e ) = T p T p + F p (3) S pecif icity ( S pc ) = T n T n + F p (4) F 1 − scor e ( F 1) = 2 ∗ P r e ∗ S en P r e + S en (5) Recent research suggests confusion matrix analysis is ef fecti v e for model v alidation [8] due to its ability to characterize data linkages and distrib utions. Information on illustrating cate gorization models is included. 4.4. Multi-classication RNN Diabetic framew ork r esults W e or g anized an e xtensi v e e xperiment, and we obtained dif ferent outputs corresponding to the perfor - mance metrics to disco v er the signicant ndings. Therefore, we illustrated the results in T able 4 for the three models. The table sho ws classier model performance in accurac y (Acc), sens iti vity (Sen), specicity (Spe), precision (Pre), F1-score (F1), and A UC. Most people use the confusion matrix to e v aluate cate gorization er - rors. W e report loss, A UC, precision, r ecall, and accurac y between training and v alidation with 300 epochs in all proposal models. T able 4. The performance metrics for the three proposed models for our RNN Diabetic frame w ork Model Acc Sen Spe P re F1 A UC RNN+ResNet152V2 97.3 96.7 95.9 96.7 96.7 97.4 GR U+ResNet152V2 98.3 97.9 98.1 98.3 98.3 99 Bi-GR U+ResNet152V2 99.8 98.1 98.6 99.8 99.8 99.4 RNN Diabetic fr ame work for identifying diabetic e ye diseases (Arwa Albelaihi) Evaluation Warning : The document was created with Spire.PDF for Python.
1838 ❒ ISSN: 2502-4752 4.4.1. Results of stage 2 model 1 (RNN+ResNet152V2 model) Figure 6 displays the ResNet152V2 model confusion matrix of the RNN+ResNet152V2 model. The gure sho ws the ef fecti v e classication of four patient statuses: DR, DME, glaucoma, and cataract, with DR images ha ving the greatest ratio, follo wed by cataract, DME, and glaucoma. Figure 6. Confusion matrix of stage 2-model 1: RNN+ResNet152V2 model 4.4.2. Results of stage 2 model 2 (GR U+ResNet152V2 model) The GR U+ResNet152V2 model demonstrated impro v ed performance with GR U, as seen in T able 4; assumed v alues were higher . In Figure 7, the confusion matrix of the GR U+ResNet152V2 model classies four patient statuses: DR, DME, glaucoma, and cataract. The greatest ratio is for DME and DR pictures, follo wed by cataract (0.9839) and glaucoma (0.9542). Figure 7. Confusion matrix of stage 2-model 2: GR U+ResNet152V2 model 4.4.3. Results of stage 2 model 2 (Bi-GR U+ResNet152V2 model) Figure 8 displays the performance indicators of the Bi-GR U+ResNet152V2 model, including loss, A UC, precision, recall, and accurac y throughout training and v alidation, along with the number of epochs. See Figure 9 for the confusion matrix of the Bi-GR U+ResNet152V2 model. Figure sho ws four precisely recog- nized patient conditions: cataract, DR, DME, and glaucoma. The DR pictures ha v e the highest ratio (0.9896), follo wed by the DME, cataract, and glaucoma. The y outperform the GR U+ResNet152V2 model b ut are infe- rior to the RNN+ResNet152V2 model. Indonesian J Elec Eng & Comp Sci, V ol. 37, No. 3, March 2025: 1830–1844 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 1839 Figure 8. Loss, A UC, precision, recall, and accurac y between the training and v alidation phases, with number of epochs for stage 2-model 3: Bi-GR U+ResNet152V2 model Figure 9. Confusion matrix of stage 2-model 3: Bi-GR U+ResNet152V2 model 4.5. Comparati v e analysis and discussion This study recommends three deep RNN models for diabetic e ye disease diagnosis: RNN+ResNet152V2, GR U+ResNet152V2, and Bi-GR U+Res Net152V2. DR, DME, cataract, and glaucoma the four most common diabetic e ye diseases are classied using these models. The proposed models were e v aluated using accu- rac y , sensiti vity , F1-score, and A UC. According to our data, Bi-GR U+ResNet152V2 had the highest clas- sication accurac y (99.8), follo wed by GR U+ResNet152V2. Ho we v er , T able 4 and Figure 10 indicate that RNN+ResNet152V2 has the smallest accurac y among all designs. T able 5 compares the performance e v aluation measures of our proposed RNN Diabetic frame w ork to e xisting DL model s from past publications. This research e v aluated the proposed model usi n g A UC, sensiti vity , specicity , precision, F1-score, and accurac y . W e used tw o gures to e xplain the comparisons. Figure 11 RNN Diabetic fr ame work for identifying diabetic e ye diseases (Arwa Albelaihi) Evaluation Warning : The document was created with Spire.PDF for Python.