Automated Staging and Grading for Retinopathy of Prematurity on Indian Database

s:Retinopathy of prematurity (ROP) is a disorder of the retina in neonates. If ROP is not treated at early stage, neonates’ vision is affected, leading to blindness. It is necessary to diagnose and treat ROP at earliest. Several ROP assessment techniques based on Image analysis have been introduced in recent years. These studies identify only normal, abnormal and plus disease. This research article explores the identification of distinct ROP stages along with normal and abnormal detection. Detecting the stages will help to expedite the treatment and prevent vision loss. The proposed framework consists of feature extraction using Scale Invariant Feature Transform (SIFT) and Pyramid Histogram of Words (PHOW) techniques. Three efficient supervised machine learning algorithms, namely random forest (RF), support vector machine (SVM) and extreme boosting gradient (XGBoost), are used to classify different stages of ROP. A data set captured by RetCam 3 is used to evaluate the model. Based on rigorous evaluation, the accuracy of different ROP stages is 93.68%, 83.33%, 85.71%, 55.55% and 100% for normal, stage 1, 2, 3 and 4, respectively.


Introduction
Retinopathy of prematurity (ROP) is a vascular disease that may cause vision impairment [1].ROP is a challenging dataset consists of stage 1, 2, 3 and 4 with a lot of variability in orientation, blood veins and intensity values.Retinopathy of prematurity can be treated if good screening is available.Good screening leads to early and accurate detection and early detection should allow for more successful treatment.There are around three hundred retina specialists in India of which only very few are ROP specialists.The automation of ROP classification will help the specialists to focus on the treatment.The software tool can be easily used by the technicians, paramedic staff and others to classify the stages of ROP.
ROP is underdeveloped vascularization of retina in premature babies which can lead to vision loss or blindness.Premature babies with birth weight less than 1800 g or gestational age less than 32 weeks are at risk of developing ROP.Due to the incomplete development of the premature baby, a vascular delay or a vascular lesion determines the appearance of ROP.The avascular area of the retina will be delimited from the vascular area by: the absence of vessels or by a line (ROP stage I), ridge (stage II) or vascular proliferation (stage III) or a retinal detachment (stage IV or V).In advanced grades of retinopathy of prematurity there is a high risk of retinal detachment and bilateral blindness.For this reason, neonates are screened at 4-6 weeks after birth, and then every 1 or 2 weeks until the vascularization of the retina is complete.
The retina consists of multiple blood vessels.If there is any irregularity in these vessels then retinopathy is developed.Development of this disorder of retina happens either slowly or suddenly.This condition is treated by an ophthalmologist by studying the images taken from different cameras.Early treatment might lead to improvement in vision and avoid further degradation of vision.
Retinopathy disease development is variable in each individual.This condition can get normalized on its own or lead to permanent damage.Earlier ROP was called retrolental fibroplasias [2].
Various studies show that the leading cause of blindness is rise in premature babies, not only in developed but also in developing countries.During initial stages of ROP there is less impact on sight.In later stages, blood vessels become more tortuous, neovascularization occurs which is an indicator of ROP.These new vessels are delicate and may cause bleeding to impact the vision.In advanced stages, there are chances of retinal detachment that endangers the vision thereby causing blindness [3].Most of the research work is carried on retinal grading of blood vessels, tracing of vessels [4] diameter, curvature [5], length to chord [6], angle based measures [7,8,9], PCA [10], tortuosity measurement for detection of pre plus, plus disease and normal and limited work on the automation of classification of stages based on severity [11,12].The challenges faced for development of automation of classification of stages are ROP dataset is limited and extremely imbalanced dataset due to lack of awareness and expertise.Once the classification of ROP stages based on severity is done timely treatment can be provided to neonates thereby preventing blindness.There is a scope to work on identifying the presence of ROP and assessing the severity of ROP and classifying them using machine learning, which will help not only to provide early treatment but also help to reach more number of patients through telemedicine.
The rest of the paper is organized as followed.Section 2 discusses related studies.Section 3 describes proposed methodology adopted in the study whereas Section 4 discusses proposed classification.Section 5 explains about the imaging modality, dataset, experiments and outcomes/results.Section 6 summarizes the findings of the study and finally, section 7 ends with future scope of the study.

Related work
International Classification of ROP (ICROP) was published in 1984 and 1987.The ICROP document was about the treatment, observations and different phases of the disease [13,14].The ICROP document was revisited in 2005 with the addition of aggressive posterior phase and pre plus disease.Last stage of ROP results in retinal detachment and blindness [15].The four features are evaluated based on ICROP classification: Location, Severity, Extent, Presence or absence of plus disease.There are three Zones based on the location Zone I, Zone II and Zone III respectively.The identification of Zone I based on the optic disc and macula structures were calculated as per the ICROP using deep learning algorithm [16].The study demonstrated the Zone II ROP at low, moderate and high risk level.The incidence of type 1 Zone II ROP has significant decrease of severe ROP in China might be due to awareness and risk of vision loss due to ROP [17].RetCam imaging can be used for characterization of retinal hemorrhage in neonatal babies, found mostly in Zone II with grades II and III.During the process, there are minor systemic symptoms that may be recovered on their own indicating the safety and efficacy of the imaging tool [18].The assessment of Retinal Lesions found in premature infants and its implication are crucial in the clinical practice of ophthalmology [19].The paper addresses the regression of stage 3 ROP and the growth of blood vessels in Zone III of the retina, which can provide valuable information for healthcare practitioners when making decisions in borderline ROP cases [20].
The artificial neural network is used to overcome the non-uniform distribution and low quality of retinal fundus images for the better visibility by improving the four features contrast, brightness, gamma factor and cliplimit of retinal blood vessels to support the ROP diagnosis [21].The more advance automated network like Convolutional Neural Network is used for complicated task of grading the retinal blood vessels.The Transfer learning approach along with pre trained Inception V3 feature for attention network is adopted for automatic diagnosis of vessels [22].The automated quality evaluations of images were classified as acceptable, possibly acceptable and not acceptable using deep convolutional neural network to detect the presence of ROP confidently [23].
A technique on a curvature based estimation algorithm for automatic evaluation of tortuosity in retinal image was addressed [5].The author proposes the multiple instances learning method for detection of ROP [24].The author has discussed vessel segmentation using COSFIRE and evaluated tortuosity using curvature method [25].The matched filter plus kernel performance was better compared to the matched filter and first-order derivative of Gaussian [26].A new approach to measure the tortuosity based on principal component analysis was proposed [10].
Author assessed vessel tortuosity using Gabor filter and Morphological methods [7].
Another set of studies discuss image localization, feature extraction and classification using Neural Network and could classify only 3 stages: stage 1, stage 2 and stage 3 respectively [11].They could establish use of segmentation of vessels to develop a supervised classification for ROP.The extracted feature set comprises multi-scales vesselness and texture features.Describes about Vessel Segmentation and accuracy of 97% is achieved [27].Manual segmentation of the vessels and extraction of image-based features, such as tortuosity and integrated curvature are discussed where the accuracy was 90% [28].
Author used vessel centerline extraction for vessel tracing and achieved sensitivity of 0.78 with 0.15 false detection rate [4].For authors, a major challenge was of non-agreement of experts while diagnosing the tortuosity of ROP than with the metrics [29].For a detailed review on applications of image processing to diagnosing ROP and comparison of different methods, is discussed [30].Author not only analyzed width but also arteriolar and venular tortuosity in advancing ROP.They concluded tortuosity makes the difference and not the width of vessels for screening of ROP [31].An algorithm was developed for measuring tortuosity of vessels but they used images with manual tracing of blood vessels [32].Author extracted distinctive invariant features of images with a method development that is named as Scale Invariant Feature Transform [33].
Artificial intelligence algorithms aid in the clinical practice of ROP for improving the accuracy and efficiency of the ROP with proper handling of image acquisition, feasibility and validation.Real world challenges in implementation, development and strategies of artificial intelligence were discussed to bring the technology for prevention of blindness due to the ROP [34].In rural areas there is a scarcity of ophthalmologists and unavailability of specialized facilities so telemedicine will aid in the screening of ROP.The geometric features like area and diameter were extracted first using Hessian analysis and then SVM classifier is used for detection and classification of ROP images as normal, stage 2 and stage 3 [12].
The detection of the first three stages and plus disease done with the help of Back Propagation Network (BPN) and a combined model of BPN and Radial Basis Function (RBF) network [35].
Moreover, in the diagnosis of diseases, computerized image processing technologies not only help to save time and reproducibility but also often more accurate than human experts [36].Transfer learning framework has been discussed for screening of ROP.There are very few ophthalmologists available for the screening of ROP.In addition, there is a shortage of staff, medical equipment and policy for ROP screening [37].Author has discussed the calculation of AVR by segmenting the vessel center line into arteries and veins and then classifying them [38].The author has proposed the convolution neural network for classification of severity of ROP and also compared mean aggregate operator.Moreover the author concluded that the illumination has an impact on image recognition and result [39].A novel approach was introduced by the author using a matched filter.The author presented a semi-automated CAD method for plus disease assessment [40].Vessel Segmentation which is one of the important features is extracted using Convolutional Neural Networks [41].The optimization framework is adopted to address the significant challenges and early detection of ROP using Convolutional Neural Networks with perfect sensitivity score [42].Different Deep Learning Models [43,44] are employed for identification of ROP, NOROP, mild severe, and Plus disease.Article [45] provides a detailed review of several deep learning algorithms have been employed in recent years for the detection of various eye disorders.

Proposed Methodology
The flow chart of the proposed methodology is shown in Fig. 2. RetCam imaging system is used for ROP Screening.Once the images are acquired pre-processing is done for further enhancement and noise removal.The feature extraction and summarization process is comprised of the following techniques: 1. Pyramid histogram of words (PHOW) features (dense multi-scale SIFT descriptors) computation on images.2. K-means for visual word dictionary construction i.e.Bag of words (BoW) on PHOW features.Training and evaluation is done using various machine based classifiers.Three different classifiers are used and finally 4 stages are classified.

Fig. 2. Block diagram of proposed workflow.
In this proposal we developed a novel framework which consists of successive steps:

A. Preprocessing
Pre-processing consists of resizing, contrast enhancement, illumination equalization and noise removal.The images are resized to 512x512x3 in the proposed work.Then the RGB images are converted to greyscale.

B. Feature Extraction: SIFT
The SIFT detects and describes local features in images.Reference images are created with extraction of the keypoints.These sets of reference images are stored in a database.The important stages of computation for SIFT are scale space, LoG approximations, finding keypoints, getting rid of low contrast keypoints, keypoint orientations, generating a feature.Building the scale space, all the scales must be examined to identify scale invariant features.The effective way to get the best scale is to compute the laplacian pyramid.Approximations of LoG are done by difference of Gaussians.The value of sigma and k has to be decided.Comparison of pixels to its 26 neighbors [33].
Each octave has 3 layers [33].The initial value of sigma is 1.6 and the numbers of octaves are computed automatically from the image resolution.Weak features from low contrast regions were filtered out by using the contrast threshold.The sigma of the Gaussian is applied to the input image at the octave.
We looked at each point of 3X3 neighborhoods in the lower and higher scale as shown in Fig. 3 and compared a pixel X with 26 pixels in current and adjacent scale .We selected the pixel X if it is larger/smaller than all 26 pixels, so that large numbers of extrema are found.Computationally it is expensive but detects the most stable subset with coarse sampling scales.Key point localization is to localize each interest point.The use of Taylor series expression of difference of Gaussian (DOG), helps to avoid the outliers.Further outliers can be removed by computing the principal curvature.Now orientation of interest point is to be assigned at each of the point or direction and then histogram of 36 bins is built as 360 degrees will give 36 bins each with 10 degrees.Sensitivity to changes in image scale of Harris detector doesn't allow its use for matching images of different sizes.

Proposed Classification
We have introduced a fully automatic algorithm in this paper for classification of different ROP stages on the basis of severity using three different classifiers.
Classification of ROP is a multiclass classification problem.There are different stages of ROP.We have considered four stages for experimentation.It is a four class classification problem of ROP, which will classify as stage 1, 2, 3 and 4. In multi class classification each image is assigned with one label.There are five stages of ROP tabulated in Table 1.
We could classify the different stages of ROP on the basis of the severity.We conducted four experiments for classifying the stages of ROP which is a multiclass problem.The individual stages were graded using different classifiers.Out of 300 images, 47 were labelled as class 1 (stage 1), 47 were labelled as class 2 (stage 2), 60 were labelled as class 3 (stage 3), 38 were labelled as class 4 (stage 4) and 108 were labelled as class 5 (normal).[48].
XGBoost is a machine learning algorithm that has recently been dominating for structured or tabular data.XGBoost is an implementation of gradient boosted decision trees designed for speed and performance [49].

Experimentation and Results
In this section, we have presented the experimental setup, imaging modality, analysis strategy and execution of the proposed method.

A. Image Acquisition
RetCam is an alternative digital imaging tool for screening ROP conditions of neonatal [50].The latest advancement of ultra-widefield imaging technologies and their clinical applications, has accelerated the focus on diabetic retinopathy, retinal vein occlusion, uveitis, and pediatric retina [51].
The ophthalmic images taken from RetCam can be stored, retrieved and used for further observation and analysis using a computer/laptop on board.The Indian dataset had 300 images 192 diseased and 108 Normal.The outcomes are then tabulated.Our dataset consists of different scale, viewpoint, rotation, illumination and orientation.It was a challenge to extract the features and match these images.To overcome this challenge there is a need of distinctive, invariant features extraction that should be reliable for matching.
The Table 2 shows the distribution of data which is unbalanced in the ROP screening.The ratio of ROP to Non ROP is 1:2.The diseased images are approximately half of the non ROP images.The RGB images of the size 1600x1200x3 were resized into 512x512x3 to save the computational burden.The resized images were converted into grayscale.All four stages dataset was given as input.We divided dataset into training, validation and testing images.The SIFT features were extracted from training: We select ROP stage 1, 2, 3 and 4 images, namely the 300 images.We compute dense SIFT descriptors and quantize them using a visual dictionary (1000 words).This collection of a dictionary is used as training data for a parameter-tuned SVM classifier with a linear kernel, Random forest and XGBoost.Finally, these trained images are given to the classifier which distinguish the images based on severity of ROP i.e. stages 1, 2, 3, and 4.

C. MINIBATCH K MEANS AND PHOW DESCRIPTORS
One of the most popular clustering algorithms is k-means but computational cost is high as the dataset increases.We have used mini batch kmeans which is an alternative to k-means.It reduces the computational cost for finding the partition of clusters.All the features similar to each other will fall into one cluster.There will be four clusters for four stages and as clusters increase, the computational cost is more evident but at the cost of quality loss of cluster.
Bag of Words (BOW) is a feature summarization technique that helps in image classification.Keypoints are extracted using SIFT for each of the images and then bag of words are created.The PHOW is used for image description which is an extension to the BOW.The PHOW helps to overcome limitation of BOW [54,55] where the information of spatial image features is available by partitioning the image into pyramids (sub regions) and the histogram of each pyramid is concatenated to the histogram of the original image with a suitable weight.).The recall is beneficial than precision in most of the cases.As the data distribution is unequal; F1 score or PR will give more insights about the result and be beneficial.
The validation accuracy of Normal, stage 1, stage 2, stage 3 and stage 4 for SVM is 89%,82%, 81% 65% and 80% whereas for RF and XGB is 88%,78%, 69%, 65% and 86% and for XGB is 92%, 78%, 69%, 59% and 86% respectively.Similarly the testing accuracy is 92%, 83%, 86%, 44% and 100% for SVM, 90%, 71%, 86%, 56% and 100% for RF and 94%, 83%, 86%, 56% and 100% for XGB respectively.The figure 7 shows the comparison of all accuracy using different classifiers.The accuracy of stage 3 is less due to an unbalanced data, quality of images, data size and low contrast.The extraction of features like demarcation line, ridge and vessel was challenging due to quality of image and low contrast.For the classification of the images based on location or Zone wise needed the optic disk (OD) fovea and macula to calculate the Zone I, Zone II and Zone III.The dataset had hardly any information or features like OD and macula.Classifying the images based on the extent which requires minimum five contiguous or eight noncontiguous clock hours to describe the disease.OD, Fovea and the macula information was required to mark clock hours.It is very challenging to extract distinctive features from fundus images of ROP as the pixel intensity is low contrast.
The SVM, RF and XGB classifiers were used to evaluate the accuracy of normal Vs abnormal, stage 1, stage 2, stage 3 and stage 4. The classification accuracy is 93.33%, 90.17% and 93.68 for normal vs. abnormal, for stage 1 is 83.33%, 71.42% and 83.33%, for stage 2 is 85.71% , stage 3 is 44.44%, 55.55% and 55.55% stage 4 is 100% respectively.The model had mean validation accuracy of 77.07%, 74.23% and 72.76% respectively after training the dataset.The models' performance was evaluated on the test set.The test set was randomly selected.The Table 3 displays mean accuracy for five folds cross validation for SVM, RF and XGBoost classifier.Our state of the art model was able to classify the multiclass problem with mean testing accuracy of 78.37%, 78.17% and 81.15% using SVM, RF and XGBoost classifiers respectively.

Discussion
The proposed method results are critically analyzed.There are relatively very few findings for evaluation of ROP Normal/Abnormal along with stages identification.Our method outperforms in the classifying the different stages which are based on the severity of the disease.Table 4 shows the computational time for different classifiers.Table 5 displays the comparison of findings by different researchers for detection and classification of ROP.
SVM takes less time as compared to RF. Computational efficiency of SVM is better than RF but Precision, recall and f1 score is better for RF.
The Table 5 depicts the accuracy comparison by different researchers.As per our knowledge work done by the researchers is focused on blood vessel detection and tortuosity and limited work has been reported on the classification of stages based on the severity of the disease.The Prabakar et.al. [11] used the histogram approach using limited dataset for classification and no performance metrics were discussed in the results.Rebecca Rollins et al. [56] classified the images into three categories No ROP (NR), ROP not requiring treatment (RNT) and ROP which required treatment (RT).The Priya Rani et al. [24] and Yinsheng Zhang et al. [37] discussed the disease as Normal and ROP.The Worrall et al [57], Junjie Hu et al. [39] and Y.P Huang [43] reported the disease as Normal, mild, severe.The geometric features like area and diameter were extracted using hessian analysis for the study of Normal, Stage 2, Stage 3 by Vijayalakshmi et al. [12].The Convolutional neural network was adopted to classify positive and negative samples by Xin Guo et.al.[42].Deep fusion feature [58] approach has been used to classify different stages.Even though the cost prohibitive concern has been resolved by MII RetCam, made in India [52], early diagnosis remain the challenge due to unavailability of automation and less number of pediatric retina ophthalmologists.So the research can be carried out for the ease and automation required for ROP stage diagnosis.The ability to image entire retina and representation of 3D to 2D would help overcome the limitations in the ROP research [51].The evolution of Zone II features were analyzed and found that there was gradual decrease in the risk with time and responded well to the treatment [17].The identification of Zone I is subjective and inaccurate so DNN algorithm was developed [16].
The implications of retinal lesions are also important in the diagnosis and might overcome the inaccuracies and vison loss [19].
Our  Extraretinal fibrovascular growth or neovascularization spreads from the ridge into the vitreous is the clinical criterion for stage 3 ROP.However, ridge identification was difficult to detect in some ROP fundus images for both clinical labelling and computerized ROP staging.As a result, learning the small distinctions between them were challenging for the depth network, resulting in prediction mistakes.Despite the fact that there were few ROP data from stage 4 in the training process, ROP detection accuracy in stage 4 was quiet high.In terms of our test data, the suggested technique successfully predicts all ROP fundus images in stage 4. One probable reason is that the retina in stage 4 has a retinal detachment, which is seen in the fundus imaging.As a result, our depth network can precisely learn its effective characteristics.

ETHICS COMMITTEE APPROVAL
Approval was taken from the institutional human ethics committee (IHEC) before the study.

Fig. 3 .
Fig. 3. a ) Generation of scale space using Gaussian and Difference of Gaussian.b) Comparison of pixels to its 26 neighbors [33].

Fig. 6
Fig. 6 shows the SIFT keypoints and histogram of stage 1.The bag of words is used to quantize the local features and classify the images.The model generates the feature vectors and maps the vectors using bag of words and then generates the histogram of keypoints.PHOW descriptors are computed and then k means is used for clustering the descriptors into centroids.The PHOW is used for training multiclass classifiers.

Fig. 6 .
Fig. 6.SIFT key points and histogram of stage 1 We treat the grading of ROP as a multiclass classification.We have five classes, i.e.Normal, Stage 1, 2, 3, and 4 respectively.The imbalanced dataset has unequal class distribution.A total of 192 dataset which are ROP cases, 47 belong to class 1(stage 1), 47 belong to class 2 (stage 2) and 60 to class 3(stage 3) and 38 belong to class four (stage 4) and 108 belong to class 5(Normal) which is Non ROP.

Fig. 8 , 9 ,
Fig. 8, 9, and 10 shows the precision; recall and f1 score for different classifiers respectively.RF classifier has high precision, recall and f1 score as compared to SVM and XGBoost on the testing data.SVM classifier has high precision, recall and f1 score on the validation data.But all the three metrics are very close to each other on the testing data set.The model has 77%, 79% and 78% precision for three classifiers respectively.The RF predicts the ROP 79% of the time it is correct.It has a small number of false positives of 21%.The precision helps to understand the classifier's ability not to misguide negative as positive.The recall graph indicates high recall for RF classifier i.e. the small number of false negatives.The model has 74.52%, 74.56% and 73.55% recall for three classifiers respectively.The graph shows high recall and high precision but they are actually very close to each other.So there are a small number of false positives and false negatives respectively.The F1 score is the balance between the precision and recall.The simple harmonic mean of Precision and recall is known as f1 score.The model has 73.5%, 74.48% and 74% f1 score for three classifiers respectively.

Fig. 8 .
Fig. 8.The graph of Precision Vs Classifiers for weighted average for validation and testing accuracy.

Fig. 9 .
Fig. 9.The graph of Recall Vs Classifiers for weighted average for validation and testing accuracy.

Fig. 10 .
Fig. 10.The graph of f1 score Vs Classifiers for weighted average for validation and testing accuracy.

Table 1 .
[46,47]y based ROP stages Support vector machine (SVM) is kernel based classifier separating the hyper plane.SVM is a binary classification algorithm and is robust to variance and small sample data.The features are scaled up by using different kernel functions.This article has been accepted for publication in a future issue of this journal, but it is not yet the definitive version.Content may undergo additional copyediting, typesetting and review before the final publication.Citation information: S S Kadge, S. L. Nalbalwar, A B Nandgaonkar, Parag Shah, V Narendran, Automated Staging and Grading for Retinopathy of Prematurity on Indian Database, Journal of Artificial Intelligence and Technology (2023), DOI: https://doi.org/10.37965/jait.2023.0235Adjustableparametersigmaplays a major role in the performance of the RBF kernel and should be carefully tuned to the problem at hand.The multiclass SVM has two categories one-versus-one (OVO) or one-versus-all (OVA)[46,47].
[53]laptop monitor is used for viewing images with the RetCam Shuttle and RetCam Portable.Proprietary software is installed on the computers to capture, store, view, retrieve, and export ophthalmic images.The key factors of the device are photo documentation, serial imaging, portability and financial affordability[52].The clinically challenging features of neonatal fundus images were enhanced with the aid of inexpensive and noninvasive RetiView" software specifically in the subset of Aggressive Posterior Retinopathy of Prematurity (APROP) images[53].

Table 2 :
Dataset used for Training, Testing and Validation This article has been accepted for publication in a future issue of this journal, but it is not yet the definitive version.Content may undergo additional copyediting, typesetting and review before the final publication.Citation information: S S Kadge, S. L. Nalbalwar, A B Nandgaonkar, Parag Shah, V Narendran, Automated Staging and Grading for Retinopathy of Prematurity on Indian Database, Journal of Artificial Intelligence and Technology (2023), DOI: https://doi.org/10.37965/jait.2023.0235

Table 3 :
Mean Accuracy and Standard deviation (SD) for SVM, RF and XGBoost classifiers Our findings provides the assessment of Normal/Abnormal ROP and classification of different stages, i.e.Stage 1, Stage 2, Stage 3 and Stage 4 over the Indian databases using SVM, RF and XGBoost classifiers.The computational time, comparison of results and limitations are also discussed here.This article has been accepted for publication in a future issue of this journal, but it is not yet the definitive version.Content may undergo additional copyediting, typesetting and review before the final publication.Citation information: S S Kadge, S. L. Nalbalwar, A B Nandgaonkar, Parag Shah, V Narendran, Automated Staging and Grading for Retinopathy of Prematurity on Indian Database, Journal of Artificial Intelligence and Technology (2023), DOI: https://doi.org/10.37965/jait.2023.0235

Table 4 .
paper not only focuses on Normal and ROP disease but also on the classification of Stage 1, Stage 2, Stage 3 and Stage 4 on the basis of severity if ROP is present.Our proposed work has fourfold advantage.Firstly we are able to classify whether ROP is present or absent.Computational Time for SVM, RF and XGBoost classifiers

Table 5 :
Comparison of accuracy for detection and classification of ROPThis article has been accepted for publication in a future issue of this journal, but it is not yet the definitive version.Content may undergo additional copyediting, typesetting and review before the final publication.Citation information: S S Kadge, S. L. Nalbalwar, A B Nandgaonkar, Parag Shah, V Narendran, Automated Staging and Grading for Retinopathy of Prematurity on Indian Database, Journal of Artificial Intelligence and Technology (2023), DOI: https://doi.org/10.37965/jait.2023.0235This research proposed a simple and efficient framework for unique ROP staging based on SIFT and three separate machine learning classifiers.In comparison to previous state-of-the-art classification approaches, the suggested method employed SIFT features which focused on lesion-related regions of ROP and effectively increase ROP staging accuracy and model generalization capacity.The results revealed that classification may considerably increase ROP staging performance and classification accuracy.Our technique tended to misclassify misclassified samples into the adjacent next severe stage of ROP image classification, which was consistent with clinical staging standards for ambiguous samples.Our suggested approach obtained an This article has been accepted for publication in a future issue of this journal, but it is not yet the definitive version.Content may undergo additional copyediting, typesetting and review before the final publication.Citation information: S S Kadge, S. L. Nalbalwar, A B Nandgaonkar, Parag Shah, V Narendran, Automated Staging and Grading for Retinopathy of Prematurity on Indian Database, Journal of Artificial Intelligence and Technology (2023), DOI: https://doi.org/10.37965/jait.2023.0235accuracy of 93.68%, which is comparable to other quantitative metrics.However, when compared to existing classification systems, our proposed method has higher recognition accuracy for stages 1, 2, and 4. The Stage 3 performance was quite low when compared to the other stages' performances.In our study, 55% of ROP fundus images with stage 3 were correctly recognized, whereas the other 45% were all incorrectly classified as stage 2 and 4.There are two basic causes behind this: (1) Scarcity of ROP data in the training phase.Although we have lessen the influence of imbalance categories by altering the weight of the loss function and employing transfer learning, we still cannot effectively fix the problem caused by data imbalance.(2)