SmartPlus: A Computer-based Image Analysis Method to Predict Continuous-valued Vascular Abnormality Index in Retinopathy of Prematurity

doi:10.21203/rs.3.rs-4171445/v1

Download PDF

Article

SmartPlus: A Computer-based Image Analysis Method to Predict Continuous-valued Vascular Abnormality Index in Retinopathy of Prematurity

https://doi.org/10.21203/rs.3.rs-4171445/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Plus disease is characterized by abnormal changes in retinal vasculature of premature infants. Presence of Plus disease is an important criterion for identifying treatment-requiring in Retinopathy of Prematurity (ROP). However, diagnosis of Plus disease has been shown to be subjective and there is wide variability in the classification of Plus disease by ROP experts, which is mainly because experts have different cut points for distinguishing the levels of vascular abnormality. This suggests that a continuous ROP Plus disease severity score may reflect more accurately the behavior of expert clinicians and may better standardize the classification. The effect of using quantitative methods and computer-based image analysis to improve the objectivity of Plus disease diagnosis have been well established. Nevertheless, the current methods are based on discrete classifications of the disease severity and lack the compatibility with the continuous nature of abnormal changes in retinal vasculatures. In this study, we developed a computer-based method that performs a quantitative analysis of vascular characteristics associated with Plus disease and utilizes them to build a regression model that outputs a continuous spectrum of Plus severity. We evaluated the proposed method against the consensus diagnosis made by four ROP experts on 76 posterior ROP images. The findings of our study indicate that our approach demonstrated a relatively acceptable level of accuracy in evaluating the severity of Plus disease, which is comparable to the diagnostic abilities of experts.

Health sciences/Health care/Medical imaging

Health sciences/Health care/Paediatrics/Retinopathy of prematurity

Retinopathy of prematurity (ROP) is a leading cause of blindness in children worldwide ¹. ROP is characterized by abnormal retinal vascular development at the boundary of vascularized and avascular peripheral retina ^2,3. The incidence of ROP in developed and developing countries is approximated to be 9% and 12% respectively⁴, however in low and middle-income countries, the risk of ROP-induced blindness is as high as 44% due to limitations in retinal examination and follow-up of patients and the availability of neonatal intensive care units ⁵. “Plus disease” describes the most severe vascular changes associated with ROP that defined by a level of vascular dilation and tortuosity.

Several studies^6–8 have shown that ROP (characterized by Plus disease) could be effectively treated with laser photocoagulation^6,7 or with intravitreal injection of bevacizumab ⁸ if the cases are diagnosed in a timely manner. Nevertheless, conventional screening method requires experienced pediatric ophthalmologists or retinal specialists that many areas in the world do not have access to such experts.

Accurate ROP diagnosis is challenging, even for infants who do undergo screening. It has been demonstrated that inter-examiner diagnostic variability is high even among experienced ROP clinicians^9–12. These factors have increased interest in artificial intelligence technologies and computer-based methods for automated detection of ROP, which could expand screening availability and improve the objectivity of ROP diagnosis.

There are numerous reasons for inter-expert variability in Plus disease diagnosis. One of the main reasons is that the experts often use different cut points for assessing the levels of vascular abnormality required for diagnosis of Plus diseased¹⁰. In fact, to model the range of vascular abnormalities in Plus disease more accurately, a continuous severity score, instead of current discrete classifications (i.e. Normal, pre-Plus, Plus) is required¹⁰. Majority of the current computer-based methods for the automated diagnosis of Plus disease are based on discrete classification of disease severity and so exhibits the mentioned limitations.

In this study, we demonstrated how breaking down the individual “Plus” level into more specific levels of severity helps clinicians to better differentiate disease severity. We also proposed a computer-based method to extract vascular features and utilize them to build a linear regression model for assessing Plus disease severity. Thus, the proposed model outputs a continuous spectrum of the disease severity score with the aim of improving the objectivity in diagnosis and facilitating a quantitative follow-up during treatments.

Ethics

Four ROP experts, conducted the collection of the images and participated in the labeling of fundus images. The present study adhered to the principles outlined in the Declaration of Helsinki and received approval from the institutional review board of Tehran University of Medical Sciences. Written informed consent was obtained from the parents of all patients involved in the study, granting permission for imaging and participation in the research.

Subjects and Reference Standard Diagnosis

We established a database of wide-angle posterior retinal images, each of which, relates to different individual preterm infants acquired during routine clinical care. Subjects were 27 female and 49 male infants who: (1) had an average birth weight of 1305 gr with a standard deviation of 427 gr, (2) had an average gestational age of 29.3 with a standard deviation of 3 weeks (3) were examined in a participating neonatal intensive care unit between January 1, 2019 and December 30, 2020 at Farabi Eye Hospital in Tehran, Iran (4) met published criteria for ROP screening examination. Using a wide-angle imaging device (RetCam; Clarity Medical Systems, Pleasanton, CA), all images were captured. The image has a resolution of 1200 by 1600 pixels. Four ROP expert observers independently categorized selected images as five levels of “Normal”, “pre-Plus”, “Plus1”, “Plus2”, and “Plus3” based on the severity of the vascular abnormalities and assign integer numerical scores of 1 through 5 respectively. A reference standard diagnosis (RSD) was defined for each of the images as the majority of the scores given by the four experts. When there were discordant scores among the experts, these were adjudicated in a session with the participation of all of the four experts to determine the standard diagnosis. To facilitate the grading process by the experts, a web-based interface was designed.

Image Selection and Analysis

Difficulties associated with retinal imaging of infants such as smaller eyes, less developed pupils, pressure on globe during image capture and poor fixation lead to low quality images in case of focusing, contrast, motion blurring, and uneven illumination artifacts^13,14. Low contrast and/or out of focus regions without clear vessel boundaries negatively affect the vessels segmentation and the accuracy of the consequent analysis. In order to address the concerns regarding image quality, an expert systematically eliminated low quality images from the dataset.

Utilizing an encoder/decoder network based on U-Net¹⁵, we automatically extracted a vessel mask for each ROP image. Regarding the mentioned artifacts in ROP images and the necessity of a very accurate vessels segmentation for the consequent analysis, we developed a graphical user interface (GUI) to correct the errors of the masks produced by the automatic segmentation. Optic discs also were segmented using the GUI. We are completing our method to obtain a high-accuracy and fully automated vessel segmentation pipeline for ROP images.

Arterial tortuosity is a key component of ROP international classification system¹⁶ for diagnosing Plus disease. Several efforts have been conducted to quantize vessels tortuosity as a main feature to discern Plus from non-Plus^17–21. We used squared-derivative-curvature²² which has been extensively used in the literature for computing retinal vessels’ tortuosity and point-based curvature of the vessels. It estimates the tortuosity as the integral of square of derivative of curvature, divided by the length of the arc of the vessel segment. This method is more accurate than the so called arc-to-cord ratio method for calculating tortuosity which fail to account for the entirety of the vessels’ network geometry²³. Moreover, using the mentioned method, we were also able to calculate point-based curvature. It is shown in recent studies that the point-based curvature of the vessels is significantly increased in Plus disease^17,18.Detailed explanation on the calculation of the curvature-based tortuosity can be found in our previous studies ^{24, 17}.

As mentioned earlier, venule dilation also could be a marker of Plus disease. In order to calculate the diameter of each vessel segment, we used the method introduced in reference ²⁴. Vessel dilation was calculated for a region extending 3-disc diameters away from the border of the optic disc (referred to as 3DD).

In a recent study¹⁷, we used algorithms for the quantification of the two main vascular characteristics related to Plus disease, namely the tortuosity (including point-based curvature) and dilation of vessels to discern Plus images from non-Plus images.

Given the proliferative nature of Plus disease, as well as vessels’ dilation and tortuosity, it is hypothesized that the density of the vessels in a retina affected by Plus disease will be higher than in a healthy retina. In previous studies, the density of retinal vessels has been extensively addressed as a potential marker of ROP^25–29, especially in relation to Plus disease^17,28. Thus, in addition to assessing tortuosity and dilatation, we also looked at the vessels' density as a potential indicator of Plus disease. The ratio of the number of pixels located on the vessels over the remaining pixels in the entire vascularized region was used to calculate the vessels density for each image.

in this study, we conducted a comparison between the mentioned five levels of Plus disease severity to determine if there are any statistically significant differences in the values of the mentioned vascular characteristics.

In order to train a linear regression model for the purpose of predicting a continuous-valued severity score of Plus, Initially, a set of 10 features from the total of 76 images in our dataset were retrieved in order to characterize the aspects of vessels’ tortuosity, including point-based curvature, dilation, and vascular density, across various regions of the retina. The areas of the retina that were examined encompassed a distance of three-disc diameters from the edge of the optic disk, commonly referred to as 3DD, as well as the entirety of the image field. To prioritize vascular features with greater discrimination power, the Neighborhood Component Analysis (NCA) technique for regression was utilized³⁰. In the end, a linear regression model was trained to predict the severity score of Plus disease using four selected features. The set of four features, denoted as F1 to F4, encompasses the following characteristics: F1) the maximum diameter of the vessels within 3DD; F2) the average tortuosity of the vessel segments in 3DD; F3) the measurement of vessel density over the entire image; F4) the average curvature value of the top 1% of curvature values in the entire image. Figure 1 depicts a schematic representation of the suggested technique, together with sample outputs observed at each step.

Inter-expert variability and the experts’ agreement

Kappa statistics were calculated for the assessment of the agreements between experts. Regarding the 5-level grading method we used, weighted kappa (linear) was calculated as a more convenient measures of agreement in our study. Table 1 summarizes kappa statistics and the absolute agreements of inter-expert variability.

The evaluation of diagnostic accuracy conducted by each expert was assessed using two distinct methods. To begin with, the process involves the computation of the level of concordance between the evaluations provided by each specialist and the Reference Standard Diagnosis (RSD). Additionally, the mean squared error between (MSE) the average score of each image (serving as a continuous-valued gold standard) and the score assigned by the expert to that particular image was computed.

Table 1. Inter-expert agreement for 5-level grading method (“Normal”, “perp-Plus”, “Plus1”, “Plus2”, and “Plus3”) expressed as weighted kappa and absolute agreements.

Grader pairs	Weighted kappa (linear)	Absolute agreement (%)
G1G2	0.63	63.2
G1G3	0.57	55.3
G1G4	0.42	32.9
G2G3	0.56	52.6
G2G4	0.50	47.4
G3G4	0.45	40.8

Table 2. Agreement for individual experts with RSD based on 5-level grading method.

Grader vs. Reference	Weighted kappa (linear)	Absolute agreement (%)
G1R	0.71	69.7
G2R	0.8	76.3
G3R	0.61	60.5
G4R	0.65	60.5

Table 3. MSE of the scores given by the experts against the average score based on 5-level grading.

Expert	MSE against average score
G1	0.1521
G2	0.2113
G3	0.2574
G4	0.4153

To each level of the severity, an integer number from 1 to 5 was assigned as follow: “Normal” =1, ‘’pre-Plus” =2, ‘’Plus1” =3, ‘’Plus2” =4, Plus3” =5. Then an average score for each image was calculated based on the assigned numbers.

To make an overall comparison between our 5-level grading method and the conventional 3-level grading method, we extracted a 3-level grading data from the our original 5-level grading data by considering grades given as either of “Plus1”, “Plus2”, or “Plus3” to an aggregated “Plus” grade. Figure 2-a and Fig. 2-b demonstrate the range of diagnoses for individual images, ordered by RSD for 5-level and 3-level grading methods respectively. Each column is associated with an individual image ranked as most severe (dark blue) to less severe (light blue) by each of the 4 experts.

Impact of vascular measures

In accordance with the preceding section, four image features F1 through F4 were used to train a regression model to predict a continuous-valued severity score. Results of comparing the different groups of images based on the mentioned features are shown in Fig. 3. To evaluate the power of each feature in discerning any pairs of severity levels, t-tests between mutual levels of severity were performed. As a complementary data to Fig. 3., Table 4 summarizes the p-values associated with each of the features in discerning each pairs of severity levels. Rows of the table includes pairs of the severity levels and each column corresponds to a feature.

Regression Analysis

Using selected features, F1 to F4, we trained a linear regression model to predict Plus severity index as a range of continuous values from 1 (less severe) to 5 (most severe). Figure 4 shows 10 sample images, their corresponding feature values, RSDs, the average scores, and the regression model’s out puts (two samples per each severity level are shown).

Table 4. Results of t-tests applied on mutual levels of Plus severity based on the four selected features. P-values corresponding to significant differences are depicted as follow: P < 0.05:*; P < 0.01:**; P < 0.001:***.

		P-values
Severity levels		Diameter	Tortuosity	Density	Curvature
'Normal'	'pre-Plus'	0.1449	0.0518	0.1059	0.0031 **
'Normal'	'Plus1'	0.0010 **	0.0293*	0.0017**	0.0001 ***
'Normal'	'Plus2'	6.0e-05 ***	0.0006 ***	5.0e-05 ***	7.9e-08***
'Normal'	'Plus3'	0.0005 ***	1.0e-05 ***	7.0e-05 ***	2.0e-05 ***
'pre-Plus'	'Plus1'	0.0046 **	0.0099 **	0.0290*	0.0002 ***
'pre-Plus'	'Plus2'	4.0e-05 ***	1.6e-06***	2.0e-05 ***	9.0e-08 ***
'pre-Plus'	'Plus3'	0.0017 **	6.06e-08 ***	4.7-e06 ***	0.0003 ***
'Plus1'	'Plus2'	0.1438	0.0262 *	0.0126 *	0.3502
'Plus1'	'Plus3'	0.0783	0.1331	6.0e-05 ***	0.3371
'Plus2'	'Plus3'	0.2825	0.7749	0.0130 *	0.4013

To assess the effectiveness of the suggested regression model, we employed an Added Variable Plot to illustrate the association between the model's output and the predictor variables. The resulting plot is depicted in Fig. 5-a. Slope of the fit line and the confidence interval bound shows the explanatory power of the regression model.

Moreover, we evaluated the accuracy of our regression model using a 5-fold cross validation and calculating MSE and Mean Absolute Error (MAE) of the model output against the average scores as a gold standard. MSEs and MAEs of each expert’s diagnosis were also calculated and compared with the MSE and MAE of the regression model respectively.

The comparison is briefly presented in Fig. 5-b. According to the data presented, MSE and MAE of the regression model are 0.23 ± 0.07 and 0.38 ± 0.05 respectively. This indicates that the model outperformed two of the experts in terms of its predictive capabilities.

The conventional screening approaches for ROP necessitate the involvement of skilled pediatric ophthalmologists or retinal specialists, which unfortunately remains inaccessible in numerous regions worldwide.

Wide-field retinal imaging, such as the RetCam, has the potential to do tele-ROP screening in conjunction with a reading center. This strategy enhanced both access to ROP screening and diagnostic objectivity. Nevertheless, the clinical diagnosis of ROP has continued to be subjective, resulting in significant variability and inconsistency in diagnosis. This disagreement has been observed even among experts in ROP diagnosis, as described in previous studies^11,12. This inter-expert heterogeneity might be attributed to a variety of factors. The available evidence suggests that certain clinicians may take into account broader perspectives than what is typically depicted in a standard photograph. Additionally, these clinicians may prioritize different abnormalities in the retina, such as venous tortuosity, hemorrhage, or ridge lines. Consequently, experts may encounter challenges in promptly identifying the most essential vascular features for accurate diagnosis. Furthermore, it is noteworthy that experts often employ varying thresholds for determining the presence of vascular abnormalities necessary for diagnosing Plus disease. By acknowledging these concerns, researchers can work towards the development of more efficient diagnostic methods and objective methodologies, so enhancing the management of ROP and mitigating the risk of visual impairment in infancy.

At present, the diagnosis of clinical Plus disease relies on the establishment of certain thresholds in vascular abnormalities, which can be categorized into either a 2-level system (Plus or non-Plus) or a 3-level system (Plus, pre-Plus, or normal)¹⁶. Nevertheless, vascular changes in Plus disease, such as increasing vessel tortuosity and dilation, are continuous phenomena. This implies that Plus disease is characterized by a spectrum of vascular abnormalities, ranging from extremely normal to extremely abnormal vessels. Furthermore, ROP experts may have differing criteria for distinguishing between categories on this spectrum of disease, which may result in greater variability among and within experts¹⁰. In addition, follow-ups during treatments may become more challenging and less objective if two- or three-level classifications are utilized. In order to tackle these problems, recent research indicates the importance of using a continuous quantitative severity scale for vascular anomalies in ROP.

In this study, 76 posterior retinal images were graded by four ROP specialists using a 5-level grading method including “Normal”, “pre-Plus”, “Plus1”, “Plus2”, and “Plus3” and assigned numbers 1 to 5 to each of the levels to calculate an average score of severity for each image. The rationale for implementing a 5-level severity grading system for retinopathy of prematurity (ROP), as opposed to a 3-level system, can be comprehended by considering the following factors:

1-The existence of Plus disease in ROP signifies a continuous range of vascular abnormalities. This implies that there is considerable variation in the severity of the condition among patients, and a singular measurement may not sufficiently encompass this range of variability.

2. The implementation of a multi-level grading system has the potential to enhance the precision and dependability of diagnostic procedures. It enables a more nuanced measure of disease progression, which can be essential in deciding the best treatment strategy^11,31.

3. It was observed that eyes with a greater number of quadrants of vascular tortuosity and dilation exhibited a higher rate of progression. This observation suggests that the implementation of a multi-level grading system has the potential to yield significant insights into the probability of illness development³².

Moreover, this method of grading helped us to calculate a more accurate average score of severity as a gold standard for training a regression model compared to a 3-level grading method in which Normal, pre-Plus, and Plus levels are assigned numbers 1,2, and 3.

An additional contribution of our study involves the utilization of image processing and machine learning techniques to present a computer-based approach for predicting the severity of Plus disease. Multiple research groups have conducted investigations into the advancement of computer-based image analysis for the automated diagnosis of ROP Plus disease^18,33–42. The primary objective is to develop an automated system for Plus detection by evaluating retinal fundus images and defining the features of vessels. Conventional algorithms employ handcrafted (HC) features extraction techniques, such as assessing vascular dilation and tortuosity, to analyze retinal fundus images and distinguish between Plus level and pre-Plus/non-Plus conditions^{18–20,43–45}.

On the other hand, Deep Convolutional Neural Networks (DCNN) can learn image features from the inputs to classify labels. DCNNs have been successfully used in the automated detection of diabetic retinopathy⁴⁶, glaucoma⁴⁷, age related macular degeneration⁴⁸ and ROP^40,42,49,50. DCNNs typically requires large numbers of good quality and balanced (among labels) training samples, which can be difficult to acquire in the ROP space, due to smaller eyes, less developed pupils, pressure on globe during image acquisition and poor fixation. Moreover, features learnt by DCNNs are not transparent or explainable to clinicians. Thus, using HC features in ROP studies has been remained effective either as an standalone method^18,21,51,52 or to provide complementary information for DCNN in image classification tasks^19,53.

The existing computer-based methods for automated diagnosis of Plus disease rely on discrete classification models, which do not adequately account for the continuous nature of abnormal changes in retinal vasculature. In contrast, as previously stated, in order to more accurately represent the characteristics of vascular abnormalities and the diagnostic practices of specialists in real-world scenarios, it is necessary to employ a continuous spectrum of severity scores rather than the existing discrete classifications of Plus, pre-Plus, and Normal.

In this study, we proposed a method that extract four image features related to the vascular characteristics i.e. tortuosity, density, curvature and diameter. Mutual t-tests among the 5 severity levels showed that the extracted features were able to significantly differentiate between the mutual levels of the disease severity in most of the cases. All four features exhibited considerable discriminatory ability between pre-Plus and Plus1 levels, a critical factor in guiding decisions regarding clinical interventions. The selected features were employed in the construction of a linear regression model. The suggested model generates a continuous scale of disease severity scores. The goal of our suggested model was to get a computer-assisted ROP diagnosis in order to promote objectivity in diagnosis and to replicate the real-world behavior of experts in identifying vascular anomalies. The findings of this investigation indicate that the proposed model exhibits a greater level of accuracy compared to two out of the four experienced experts included in the study. The model described in this study may also have potential utility in future research endeavors focused on the screening of ROP reactivation following anti-vascular endothelial growth factor (anti-VEGF) treatment.

This study has a number of drawbacks. First off, our dataset included a limited number of ROP images, which could have affected the model's performance. Second, the fundus images were acquired from a single clinical site with consistent device settings and population traits, which might have decreased the diversity of the data and impacted the algorithm's capacity to generalize to other populations. Also, the algorithm's accuracy in measuring Plus disease severity could potentially be enhanced in future studies by software design improvements that incorporate larger data sets and a greater number of expert opinions.

Due to the high quality and optimal focus of the images investigated in this study, it remains uncertain whether the algorithm exhibits comparable effectiveness when lower quality images are employed.

Variable pressures utilized with the RetCam contact camera, which can influence tortuosity and vascular diameter, can potentially affect the results of current study. Also, measurement bias can be influenced by various factors such as the mode and timing of imaging, the imaging angle, and the administration of mydriatics.

In conclusion, the proposed method in this study provides a continuous spectrum severity based on vascular abnormalities with high accuracy in detecting Plus severity score, performing similarly to experts’ diagnoses. By objectively analyzing vessel characteristics, it is possible to quantitatively assess the features of disease progression. The automated system has the potential to improve physicians' ability to diagnose Plus disease, making contributions to the management of ROP by image-based telemedicine methods.

Author Contribution

S.-M.S.: designed, conceptualized, and drafted the manuscript for content, designed and implemented the algorithms, and analyzed the data. N.E.: participated in the initial diagnosis, participated in image grading and revised the manuscript for content; R.R: participated in the initial diagnosis and revised the manuscript for content; A.D.-F.: participated in the initial diagnosis, participated in image grading and revised the manuscript for content; M.I.-F.: participated in the initial diagnosis, participated in image grading and revised the manuscript for content; G.G: participated in the initial diagnosis and revised the manuscript for content; E.K.-P.: designed, conceptualized, and drafted the manuscript for content, made initial diagnosis, supervised the images acquisition, participated in image selection and grading, and supervised the study.

Data Availability

The datasets developed and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Gilbert C. Retinopathy of prematurity: A global perspective of the epidemics, population of babies at risk and implications for control. Early Hum Dev. 2008;84:77–82.
Stahl A, Göpel W. Screening and Treatment in Retinopathy of Prematurity. Dtsch Arzteblatt Int. 2015;112:730–735.
Sommer A, Taylor HR, Ravilla TD, et al. Challenges of Ophthalmic Care in the Developing World. JAMA Ophthalmol. 2014;132:640–644.
Chang JW. Risk factor analysis for the development and progression of retinopathy of prematurity. PLoS ONE. 2019;14:e0219934.
Senjam SS, Chandra P. Retinopathy of prematurity: Addressing the emerging burden in developing countries. J Fam Med Prim Care. 2020;9:2600–2605.
Multicenter Trial of Cryotherapy for Retinopathy of Prematurity: Preliminary Results. Arch Ophthalmol. 1988;106:471.
Good WV, Early Treatment for Retinopathy of Prematurity Cooperative Group. Final results of the Early Treatment for Retinopathy of Prematurity (ETROP) randomized trial. Trans Am Ophthalmol Soc. 2004;102:233–248; discussion 248–250.
Mintz-Hittner HA, Kennedy KA, Chuang AZ, BEAT-ROP Cooperative Group. Efficacy of intravitreal bevacizumab for stage 3 + retinopathy of prematurity. N Engl J Med. 2011;364:603–615.
Chiang MF, Jiang L, Gelman R, et al. Interexpert agreement of plus disease diagnosis in retinopathy of prematurity. Arch Ophthalmol Chic Ill 1960. 2007;125:875–880.
Campbell JP, Kalpathy-Cramer J, Erdogmus D, et al. Plus Disease in Retinopathy of Prematurity: A Continuous Spectrum of Vascular Abnormality as a Basis of Diagnostic Variability. Ophthalmology. 2016;123:2338–2344.
Kalpathy-Cramer J, Campbell JP, Erdogmus D, et al. Plus Disease in Retinopathy of Prematurity: Improving Diagnosis by Ranking Disease Severity and Using Quantitative Image Analysis. Ophthalmology. 2016;123:2345–2351.
Bolón-Canedo V, Ataer-Cansizoglu E, Erdogmus D, et al. Dealing with inter-expert variability in Retinopathy of Prematurity: a machine learning approach. Comput Methods Programs Biomed. 2015;122:1.
Valikodath N, Cole E, Chiang MF, et al. Imaging in Retinopathy of Prematurity. Asia-Pac J Ophthalmol Phila Pa. 2019;8:178–186.
Agarwal K, Vinekar A, Chandra P, et al. Imaging the pediatric retina: An overview. Indian J Ophthalmol. 2021;69:812.
Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF, eds. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Lecture Notes in Computer Science. Cham: Springer International Publishing; 2015:234–241.
International Committee for the Classification of Retinopathy of Prematurity. The International Classification of Retinopathy of Prematurity revisited. Arch Ophthalmol Chic Ill 1960. 2005;123:991–999.
Sharafi SM, Ebrahimiadib N, Roohipourmoallai R, et al. Automated diagnosis of plus disease in retinopathy of prematurity using quantification of vessels characteristics. Sci Rep. 2024;14:6375.
Ataer-Cansizoglu E, Bolon-Canedo V, Campbell JP, et al. Computer-Based Image Analysis for Plus Disease Diagnosis in Retinopathy of Prematurity: Performance of the “i-ROP” System and Image Features Associated With Expert Diagnosis. Transl Vis Sci Technol. 2015;4:5.
Wilson CM, Cocker KD, Moseley MJ, et al. Computerized analysis of retinal vessel width and tortuosity in premature infants. Invest Ophthalmol Vis Sci. 2008;49:3577–3585.
Pour EK, Pourreza H, Zamani KA, et al. Retinopathy of Prematurity-assist: Novel Software for Detecting Plus Disease. Korean J Ophthalmol KJO. 2017;31:524–532.
Keck KM, Kalpathy-Cramer J, Ataer-Cansizoglu E, et al. Plus disease diagnosis in retinopathy of prematurity: vascular tortuosity as a function of distance from optic disc. Retina Phila Pa. 2013;33:1700–1707.
Patasius, M., Marozas, V., Lukosevicius, A. & Jegelevicius, D. Evaluation of tortuosity of eye blood vessels using the integral of square of derivative of curvature. In:; 2005.
Kashyap V, Gharleghi R, Li DD, et al. Accuracy of vascular tortuosity measures using computational modelling. Sci Rep. 2022;12:865.
Sayed Mehran Sharafi. Investigation of the Retinal Biomarkers of Alzheimer’s Disease and Atherosclerosis Using Hyperspectral Images. 2019.
Li W, Zhang W, Zhang C, et al. Soluble Tei2 fusion protein inhibits retinopathy of prematurity occurrence via regulation of the Ang/Tie2 pathway. Exp Ther Med. 2019;18:614–620.
Rezar-Dreindl S, Eibenberger K, Told R, et al. Retinal vessel architecture in retinopathy of prematurity and healthy controls using swept‐source optical coherence tomography angiography. Acta Ophthalmol (Copenh). 2021;99:e232.
Lee G-I, Park K-A, Oh SY, Kim SJ. Retinal and peripapillary vascular deformations in prematurely born children aged 4–12 years with a history of retinopathy of prematurity. Sci Rep. 2023;13:3130.
Mao J, Luo Y, Liu L, et al. Automated diagnosis and quantitative analysis of plus disease in retinopathy of prematurity based on deep convolutional neural networks. Acta Ophthalmol (Copenh). 2020;98:e339–e345.
A C, W H, M J, et al. Morphology and Vessel Density of the Macula in Preterm Children Using Optical Coherence Tomography Angiography. J Clin Med. 2022;11.
Yang W, Wang K, Zuo W. Neighborhood Component Feature Selection for High-Dimensional Data. J Comput. 2012;7:161–168.
Taylor S, Brown JM, Gupta K, et al. Monitoring Disease Progression With a Quantitative Severity Scale for Retinopathy of Prematurity Using Deep Learning. JAMA Ophthalmol. 2019;137:1022–1028.
Cheng QE, Quinn GE, Daniel E, et al. Progression from preplus to plus disease in the Telemedicine Approaches to Evaluating Acute-Phase Retinopathy of Prematurity (e-ROP) Study: incidence, timing, and predictors. J AAPOS Off Publ Am Assoc Pediatr Ophthalmol Strabismus. 2020;24:354.e1-354.e6.
Brown JM, Campbell JP, Beers A, et al. Automated Diagnosis of Plus Disease in Retinopathy of Prematurity Using Deep Convolutional Neural Networks. JAMA Ophthalmol. 2018;136:803.
Simkin SK, Misra SL, Han JV, et al. Auckland regional telemedicine retinopathy of prematurity screening network: A 10-year review. Clin Experiment Ophthalmol. 2019;47:1122–1130.
Dai S, Chow K, Vincent A. Efficacy of wide-field digital retinal imaging for retinopathy of prematurity screening. Clin Experiment Ophthalmol. 2011;39:23–29.
Murakami Y, Jain A, Silva RA, et al. Stanford University Network for Diagnosis of Retinopathy of Prematurity (SUNDROP): 12-month experience with telemedicine screening. Br J Ophthalmol. 2008;92:1456–1460.
Shah PK, Ramya A, Narendran V. Telemedicine for ROP. Asia-Pac J Ophthalmol Phila Pa. 2018;7:52–55.
Richter GM, Williams SL, Starren J, et al. Telemedicine for retinopathy of prematurity diagnosis: evaluation and challenges. Surv Ophthalmol. 2009;54:671–685.
Ells AL, Holmes JM, Astle WF, et al. Telemedicine approach to screening for severe retinopathy of prematurity: a pilot study. Ophthalmology. 2003;110:2113–2117.
Tan Z, Simkin S, Lai C, Dai S. Deep Learning Algorithm for Automated Diagnosis of Retinopathy of Prematurity Plus Disease. Transl Vis Sci Technol. 2019;8:23.
Redd TK, Campbell JP, Brown JM, et al. Evaluation of a deep learning image assessment system for detecting severe retinopathy of prematurity. Br J Ophthalmol. 2019;103:580–584.
Wang J, Ju R, Chen Y, et al. Automated retinopathy of prematurity screening using deep neural networks. EBioMedicine. 2018;35:361–368.
Binenbaum G. Algorithms for the Prediction of Retinopathy of Prematurity based upon Postnatal Weight Gain. Clin Perinatol. 2013;40:261–270.
Wittenberg LA, Jonsson NJ, Chan RVP, Chiang MF. Computer-based image analysis for plus disease diagnosis in retinopathy of prematurity. J Pediatr Ophthalmol Strabismus. 2012;49:11–19; quiz 10, 20.
E.R R, Rani P, Rajamani K, et al. Multiple Instance Learning: Robust Validation on Retinopathy of Prematurity. Int J Control Theory Appl. 2016;9.
Grzybowski A, Brona P, Lim G, et al. Artificial intelligence for diabetic retinopathy screening: a review. Eye. 2020;34:451–460.
Devalla SK, Liang Z, Pham TH, et al. Glaucoma management in the era of artificial intelligence. Br J Ophthalmol. 2020;104:301–311.
Yan Q, Weeks DE, Xin H, et al. Deep-learning-based Prediction of Late Age-Related Macular Degeneration Progression. Nat Mach Intell. 2020;2:141–150.
Fulton AB, Hansen RM, Moskowitz A, Akula JD. The neurovascular retina in retinopathy of prematurity. Prog Retin Eye Res. 2009;28:452–482.
Tong Y, Lu W, Deng Q, et al. Automated identification of retinopathy of prematurity by image-based deep learning. Eye Vis. 2020;7:40.
Pour EK, Pourreza H, Zamani KA, et al. Retinopathy of Prematurity-assist: Novel Software for Detecting Plus Disease. Korean J Ophthalmol KJO. 2017;31:524–532.
Jemshi KM, Sreelekha G, Sathidevi PS, et al. Plus disease classification in Retinopathy of Prematurity using transform based features. Multimed Tools Appl. 2023.
Yildiz VM, Tian P, Yildiz I, et al. Plus Disease in Retinopathy of Prematurity: Convolutional Neural Network Performance Using a Combined Neural Network and Feature Extraction Approach. Transl Vis Sci Technol. 2020;9:10.

No competing interests reported.

Download PDF

Reviews received at journal
12 Jul, 2024
Reviewers agreed at journal
30 Jun, 2024
Reviewers invited by journal
02 May, 2024
Editor assigned by journal
11 Apr, 2024
Editor invited by journal
10 Apr, 2024
Submission checks completed at journal
10 Apr, 2024
First submitted to journal
26 Mar, 2024

You are reading this latest preprint version

SmartPlus: A Computer-based Image Analysis Method to Predict Continuous-valued Vascular Abnormality Index in Retinopathy of Prematurity

Status:

Version 1

Abstract

Figures

Introduction

Methods

Ethics

Subjects and Reference Standard Diagnosis

Image Selection and Analysis

Results

Inter-expert variability and the experts’ agreement

Regression Analysis

Discussion

Declarations

Author Contribution

Data Availability

References

Additional Declarations

Status:

Version 1