Automated detection of type 1 ROP, type 2 ROP and A-ROP based on deep learning

Retinopathy of prematurity (ROP) is a vision-threatening vasoproliferative retinal disease in premature infants with incomplete retinal vascular development. With improving neonatal survival, the burden of screening and the need for early diagnosis and treatment are increasing. Clinical diagnosis is standardized by the International Classification of Retinopathy of Prematurity (ICROP), with disease zone, stage, and presence of pre-plus/plus remaining central to diagnosis. While most infants present with early stages (1–2) that regress spontaneously, about 10% progress to Type 1 ROP requiring treatment. Aggressive ROP (A-ROP) is a severe form defined by pathological neovascularization and plus disease without typical stage progression, risking vision loss if untreated. Increasing use of fundus photography has enabled automated image analysis and deep learning (DL) approaches, particularly convolutional neural networks (CNNs), to assist in diagnosis from complex medical images. Prior DL studies have addressed plus disease detection, general ROP detection, and vascular severity scoring. The research question of this study is whether DL-based analysis of fundus images can automatically and accurately detect and classify Type 1 ROP, Type 2 ROP, and A-ROP, potentially aiding clinical decision-making and reducing workload.

The discussion situates the work within prior DL applications for ROP. Huang et al. employed a deep CNN to distinguish early-stage (stages 1–2) ROP from no-ROP, achieving high accuracy and specificity. Wang et al. developed Id-Net and Gr-Net to detect ROP and grade severity, reporting 96% sensitivity/99% specificity for Normal vs ROP and 88%/92% for Mild vs Severe, but without staging, plus disease assessment, or ROP type classification. Brown et al. (i-ROP consortium) classified normal, pre-plus, and plus disease with 93% sensitivity/94% specificity for plus and 100%/94% for pre-plus or worse, outperforming experts; other systems like i-ROP and ROPTool also focus on plus disease. Tong et al. analyzed 36,231 images, achieving 0.93 accuracy in severity classification (ResNet-101), 0.95 in stage detection (Faster R-CNN), and 0.89 in plus detection (Faster R-CNN). Li et al. used Retina U-Net for segmentation and DenseNet for classifying normal and stages 1–3, obtaining high stage-detection accuracy; similar to the present study, they used an ROI around the demarcation/ridge to assess vascular proliferation. Redd et al. demonstrated that i-ROP DL using posterior pole images detects clinically significant ROP (Type 1, Type 2, pre-plus) with AUROC of 0.96 for Type 1, 0.91 for clinically significant ROP, and 0.99 for plus disease. Compared with these, the present work emphasizes automated discrimination among Type 1 ROP, Type 2 ROP, and A-ROP, a less commonly addressed task.

Study design and population: Retrospective study approved by the Ethics Review Board (AEŞH-EK1-2023-432) and conducted in accordance with the Declaration of Helsinki. Included were 317 premature infants meeting national screening criteria (GA < 34 weeks and BW ≤ 1700 g, or GA ≥ 34 weeks and BW > 1700 g with unstable clinical condition) with available fundus images from January 2010 to December 2022. Demographics (birth weight [BW], gestational age [GA], gender), ROP stages and zones, treatments, and postmenstrual age (PMA) at treatment were recorded. Imaging: Fundus photographs were acquired using the Heine Video Omega 2C indirect ophthalmoscope from the posterior pole (optic disc and macula) and peripheral retina. Clinical management referenced ETROP and BEAT-ROP criteria; Type 1 ROP received laser photocoagulation (LPC) or intravitreal bevacizumab (IVB) depending on zone; posterior zone II ROP also received IVB to mitigate long-term laser complications. Dataset curation: Initially 634 images from 317 infants were collected. Images of poor quality (optical artifacts, excessive light, hazy periphery, low resolution) were excluded. Final datasets: (a) ROP type classification: 189 images comprising 41 A-ROP, 56 Type 1 ROP, and 92 Type 2 ROP; (b) Stage classification from peripheral images via ROI: 43 images for stage 2 and 12 images for stage 3 (total 55). Labeling followed ICROP-3 criteria and severity groupings (Type 1, Type 2, A-ROP; stages 1–3). Image pre-processing: (1) Segmentation to isolate the lens/retinal area and remove background/noise using Adaptive Background Removal with Edge Attention; (2) Image enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE) to reduce blur; (3) Resizing to 224×224×3 to fit the classifier input. Posterior pole images were also categorized into pre-plus and plus using RegNetY002. ROI determination for peripheral images: ROP boundaries were detected with Canny edge detection; YOLOv7 then localized detected contours, producing rectangular ROIs centered on the demarcation line/ridge. Cropped ROI images were used for stage classification. Workflow logic: After pre-processing, infants with stage 1 or 2 disease in zones II–III, with or without pre-plus disease, were classified as Type 2 ROP. For others, ROI determination from temporal peripheral retina enabled classification into Type 1 ROP (plus disease with staging) or A-ROP (plus disease without typical staging). Deep learning model and training: RegNetY002 CNN served as the classifier for all tasks. Stratified 10-fold cross-validation (CV) was used, with random partitioning into 10 equal segments while ensuring homogeneity within folds. In each fold, 90% of data (n=170 for the 189-image task) were used for training and 10% (n=19) for validation; this was iterated so each subset served once as test/validation. Training parameters were fixed across experiments: 50 epochs and learning rate 0.001. Performance metrics included accuracy, specificity, ROC curves, and AUC, computed using SPSS v25 for statistical summaries.

- Dataset: 189 images for ROP type classification (41 A-ROP, 56 Type 1, 92 Type 2); 55 peripheral ROI images for stage classification (43 stage 2, 12 stage 3). - Type 2 vs (Type 1 + A-ROP): Accuracy 0.98; Specificity 0.98; AUC 0.98; 186/189 images correctly detected. - Stage 2 vs Stage 3 (ROI): Accuracy 0.90; Specificity 0.95; AUC 0.85; 50/55 images correctly detected. - A-ROP vs Type 1 ROP (ROI): Accuracy 0.91; Specificity 0.92; AUC 0.91; 96/97 images correctly detected. - Demographics: Mean GA 28 ± 2 weeks (23–34); mean BW 1156 ± 404 g (400–2770 g); 51.4% female. Treatments: 46.4% received treatment. A-ROP: 97.2% anti-VEGF, 2.8% LPC. Type 1 ROP: 27% anti-VEGF, 73% LPC. Type 2 ROP: 53.6% observed with spontaneous regression. A-ROP had lower BW and earlier PMA at treatment than Type 1 ROP (PMA 33.83 ± 1.83 vs 36.46 ± 2.46; p=0.000).

The study demonstrates that a DL pipeline using RegNetY002 on clinically acquired fundus images, with targeted pre-processing and ROI extraction around the demarcation/ridge, can accurately classify ROP into clinically meaningful categories. High accuracy, specificity, and AUC for distinguishing Type 2 from (Type 1 + A-ROP) and for discriminating A-ROP from Type 1 using peripheral ROI, indicate that combining posterior pole assessment with peripheral ridge-focused analysis captures key diagnostic features. The approach addresses a gap in prior literature, as many existing systems focus primarily on plus disease or broad Normal/ROP distinctions and do not explicitly separate Type 1, Type 2, and A-ROP. Compared with prior DL models (e.g., i-ROP, ROPTool, ResNet/Faster R-CNN approaches), the present system emphasizes classification across ROP types and demonstrates competitive performance metrics. Clinically, such automated tools could support ophthalmologists by prioritizing and standardizing assessments, potentially improving efficiency and aiding decision-making in screening and management. Nonetheless, the generalizability is constrained by dataset size and composition, indicating the need for broader validation.

DL-based analysis of fundus images can distinguish Type 1 ROP, Type 2 ROP, and A-ROP with high accuracy and specificity using a workflow that includes pre-processing, ROI detection of the ridge, and RegNetY002 classification. Integration of such AI tools into clinical practice may reduce ophthalmologists’ workload and support decision-making in ROP management.

- Retrospective design and small sample size. - Lack of advanced-stage cases (stages 4–5) may have biased dataset distribution and model performance. - Limited number of training images; exclusion of poor-quality images may affect performance estimates and limit generalizability. - Availability of fundus images influenced group distributions, contributing to a relatively high treatment rate in the cohort.