The present study was conducted by the principles of the Declaration of Helsinki following the approval of the Ethics Review Board (AEŞH-EK1-2023-432).

Three hundred seventeen premature infants who met the screening criteria according to the national screening guideline (premature infants with GA < 34 weeks and BW ≤ 1700 g or GA ≥ 34 weeks and BW > 1700 g, whose clinical condition was unstable) [10] and whose fundus images were available between January 2010-December 2022 were included in the study.

In addition to demographic information such as BW, GA, and gender, ROP stages and zones, treatment options, and postmenstrual age (PMA) at treatment were recorded. Fundus images of all premature infants taken with the Heine Video Omega® 2C indirect ophthalmoscope (Heine Optotechnik, Herrsching, Germany) were evaluated. Images were taken from the posterior pole (optic disc and macula) and peripheral retina. As stated in the ETROP study [6], laser photocoagulation (LPC) was applied to infants with Type 1 ROP and intravitreal bevacizumab (IVB) treatment was applied to infants with zone I ROP according to the results of the Bevacizumab Eliminates the Angiogenic Thread of ROP (BEAT-ROP) study [11]. Again, IVB treatment was applied to infants with posterior zone II ROP to prevent possible long-term complications that may occur after laser treatment, since the ROP line is close to the border of zone I in the posterior and has a large avascular area.

CNN architectures, which consist of multiple convolutional layers, perform operations that modify data with data-specific learning properties. In the literature, there are different architectures based on the CNN principle. While there are many layers in the designs of the architectures, the related architectures differ from each other mainly in the number of layers and the size of the new data to be produced according to the parameters in the layers. In this study, RegNetY002, which is a CNN subtype, will be used [12]. The reason why this architecture is preferred is its high performance, efficiency, scalability, and innovativeness. We used the stratified 10-fold cross-validation (CV) method during the training to evaluate and standardize our model. The samples were randomly partitioned into 10 equal-sized segments with meticulous attention paid to ensuring homogeneity and similarity within each segment. In this method, each part is set aside for testing, while the remaining parts are employed for training. This iterative process continues until every part has been utilized for testing. Thus, 90% (n = 170 images) of the data is used for training and 10% (n = 19) for validation in each trial. In this way, by obtaining a homogeneous distribution in each fold and using each subset as both training and test data, it was possible to evaluate the performance of the model independently of the subsets and to reduce the risk of memorization of the model. Additionally, to compare the performance of the model under equal conditions, the model was trained using 50 epochs and 0.001 learning rate parameters in all experiments.

To develop the algorithm, a total of 634 fundus images of 317 infants were collected. In the collected images, images with poor image quality due to optical artifacts, excessive light exposure, hazy peripheral fundus images, and low-resolution images were not evaluated. Finally, a total of 189 images, 41 for A-ROP, 56 for Type 1 ROP, and 92 for Type 2 ROP, were used to detect and classify ROP. In addition, 43 and 12 images were used to detect stage 2 and stage 3 ROP from peripheral images, respectively. Fundus images are classified and labeled according to the ICROP-3 [5] diagnostic criteria, and they are also graded and labeled according to the severity of ROP: Type 1 ROP-Type 2 ROP-A-ROP and Stage 1 ROP-Stage 2 ROP and Stage 3 ROP. To increase the classification performance of these architectures, raw images were subjected to various pre-processes.

Before analysis with DL, we processed the original images in our dataset to fit the image into the model. The pre-processing steps consisted of image enhancement, image segmentation, and resizing. First of all, the lens area in the raw images will be determined using the Adaptive Background Removal with Edge Attention algorithm, and the parts that do not contain meaningful data and have noise will be eliminated [13]. Then, Contrast Limited Adaptive Histogram Equalization (CLAHE) will be used to remove blur in the images after the segmentation process [14]. Finally, the images will be resized to 224 × 224 × 3 format to adapt to classification models (Fig. 1A). After resizing, posterior pole images were classified into two categories pre-plus and plus disease using RegNetY002 (Fig. 1C).

A Raw images were subjected to various pre-processing steps (image enhancement, image segmentation and resizing) in order to increase the classification performance of the model. B After pre-processing, the boundaries of ROP were detected in the peripheral retinal images using the Canny edge detection algorithm. Then, using the Yolo v7 detector algorithm, the detected contours were determined by red rectangles. As a result, we obtained a rectangular region (ROI). C ROP classification was made from the images obtained after pre-processing and ROI determination using RegNetY002.

After pre-processing, the boundaries of ROP were detected in the peripheral retinal images using the Canny edge detection algorithm [15]. Then, using the Yolo v7 detector algorithm, the detected contours were determined by red rectangles [16]. As a result, we obtained a rectangular region (ROI) (Fig. 1B). To study the ridge line, we made sure the rectangle was on this contour. Images were then classified according to ROP staging using RegNetY002 (Fig. 1C). Images of the ROI determination process in infants with stage 2 and stage 3 ROP are shown in Fig. 2. As shown in Fig. 2, the original retinal images are in the first column, the images obtained after pre-processing are in the second column, the images related to the ROP contours determined by the red rectangle (ROI) are in the third column, and the ROI area used to train the model is in the fourth column.

The original retinal images are in the first column, the images obtained after pre-processing are in the second column, the images related to the ROP contours determined by the red rectangle (ROI) are in the third column, and the ROI area used to train the model is in the fourth column.

As a result, after pre-processing the images, infants with stage 1 or 2 disease in zones II-III, with or without preplus disease, were classified as Type 2 ROP. Then, ROI determination was performed from peripheral retina images from the temporal side of the infants other than Type 2 ROP. In this way, infants with plus disease and ROP staging will be classified as Type 1 ROP, and infants with plus disease without ROP staging will be classified as A-ROP.

For statistical analysis, Statistical Package for the Social Sciences (SPSS Inc., Chicago, Illinois, USA) version 25.0 was used. Categorical data were expressed as numbers (n) and percentage (%), and descriptive data as mean ± standard deviation (SD). The classification performance of the model was reported as accuracy and specificity and described graphically via the receiver operating characteristic (ROC) curve. Additionally, the quantitative performance of the model was summarized by the area under the curve (AUC).