Retinopathy of prematurity (ROP) is a vision-threatening disease affecting premature infants. Early diagnosis and treatment are crucial, and the International Classification of Retinopathy of Prematurity (ICROP) provides standardized diagnostic criteria. While many infants experience early-stage ROP that regresses naturally, a significant subset develops Type 1 ROP or A-ROP, requiring intervention. The increasing use of fundus photography has led to the exploration of automated image analysis techniques, with deep learning (DL) emerging as a promising approach for improved ROP diagnosis. This study aims to develop and evaluate a DL-based model for automated detection and classification of Type 1 ROP, Type 2 ROP, and A-ROP using CNNs, analyzing fundus images obtained from clinical examinations. The potential benefits of such a model include improved diagnostic accuracy, reduced workload for ophthalmologists, and faster decision-making in the management of ROP.

Previous studies have utilized deep learning for ROP detection, focusing on tasks such as distinguishing early-stage ROP from no ROP [17], classifying ROP severity and detecting plus disease [7, 8, 21], and detecting clinically significant ROP [23]. While some studies showed high accuracy and specificity in detecting plus disease or early-stage ROP, fewer studies have addressed the classification of Type 1 ROP, Type 2 ROP, and A-ROP specifically. Existing systems like i-ROP and ROPTool primarily focus on plus disease detection [18-20]. This study aims to address this gap by developing a model capable of classifying these three distinct ROP types.

This retrospective study included fundus images from 317 premature infants (mean gestational age 28 ± 2 weeks; mean birth weight 1156 ± 404 g) who met specific screening criteria. Images were acquired using a Heine Video Omega 2C indirect ophthalmoscope. Preprocessing steps included image enhancement (adaptive background removal, CLAHE), segmentation, and resizing to 224 x 224 x 3. A region of interest (ROI) was determined in peripheral retinal images using Canny edge detection and Yolo v7. ROP classification was performed using the RegNetY002 CNN architecture. Stratified 10-fold cross-validation was employed for model training and evaluation. The model's performance was assessed using accuracy, specificity, ROC curves, and AUC scores. Images with poor image quality were excluded, resulting in a final dataset of 189 images for the primary classification task (Type 1, Type 2, A-ROP) and a smaller subset for analyzing stage 2 vs. stage 3 ROP.

The RegNetY002 model achieved high accuracy and specificity in several classification tasks. In distinguishing Type 2 ROP from Type 1 ROP and A-ROP, the model achieved 0.98 accuracy and 0.98 specificity, with an AUC of 0.98. For the analysis of ROI regions in peripheral images, the model achieved 0.90 accuracy and 0.95 specificity in detecting stage 2 ROP versus stage 3 ROP (AUC = 0.85) and 0.91 accuracy and 0.92 specificity in detecting A-ROP versus Type 1 ROP (AUC = 0.91). The model correctly classified 186 of 189 images in the primary classification task, 50 of 55 images in the stage 2 vs. stage 3 task, and 96 of 97 images in the A-ROP vs. Type 1 ROP task.

The high accuracy and specificity achieved by the DL model in classifying different types and stages of ROP suggest its potential as a valuable clinical tool. The model's performance surpasses that reported in some previous studies which often focused on a more limited set of ROP characteristics or stages. While the study demonstrates promising results, the relatively small sample size and retrospective design are limitations. The exclusion of images with poor image quality might affect the generalizability of the findings. Future research should focus on validating the model on larger, more diverse datasets, potentially including images from different imaging devices and incorporating advanced stages of ROP (stages 4-5).

This study demonstrates that DL-based analysis of fundus images can accurately classify different types and stages of ROP. The high accuracy and specificity achieved suggest that integrating this AI algorithm into clinical practice could significantly improve ROP diagnosis, reduce ophthalmologist workload, and facilitate timely intervention. Future work should focus on validating these results in larger, prospective studies and exploring the model's performance with diverse imaging modalities.

This study has several limitations, including its retrospective design and relatively small sample size, which may limit the generalizability of the findings. The exclusion of images with poor quality might have skewed the results and the high treatment rate could be attributed to the disproportionate representation of infants in the included groups due to the limited number of fundus images available. Furthermore, the absence of advanced stages of ROP (stages 4-5) might have affected the model’s performance and dataset distribution.