Research Articles Open Access

Automated Retinopathy of Prematurity Screening via Interpretable and Efficient Transfer Learning Architectures

Muhammad Haseeb ur Rehman , Saqib Dar
Available online July 1, 2026
PDF

Abstract

Retinopathy of prematurity (ROP) can progress to severe visual loss in preterm infants when treatment-requiring disease is missed. We assessed six transfer-learning CNN configurations for four-class classification of retinal fundus images: VGG19, DenseNet121, MobileNetV2, AlexNet, EfficientNetV2-S, and CBAM-ResNet50. In the last configuration, the Convolutional Block Attention Module (CBAM) is integrated with a ResNet50 backbone. The dataset comprised 2,580 images from four diagnostic categories: Healthy, Type~1 ROP, Type~2 ROP, and retinal detachment (RD), using both publicly available and clinical images. Images were preprocessed and augmented during training, and model performance was assessed with 5-fold cross-validation. CBAM-ResNet50 produced the highest reported accuracy and F1-score, reaching 95.4% and 95.9%, respectively. Its fold-specific macro-AUC ranged from 0.700 in Fold 1 to 0.975 in Fold 5, Fold 4 reached 0.942, and the mean across the five folds was 0.854. The variation between folds suggests that performance depends on the composition of the held-out data. 

Keywords

Retinopathy Machine learning Early diagnosis Transfer learning CNN

Introduction

ROP develops in the incompletely vascularized retina of preterm infants and is associated particularly with prematurity and low birth weight. Abnormal vascular growth can follow disruption of normal retinal maturation, while changes in oxygen exposure after birth may further affect this process . Advances in neonatal care have improved survival among extremely preterm infants, increasing the number of infants who require repeated ROP examinations. Providing these examinations is difficult in settings with limited access to trained ophthalmologists and organized screening services. Untreated ROP may progress to retinal detachment (RD) and permanent visual impairment. Screening therefore depends on identifying treatment-requiring disease before advanced retinal damage occurs. Clinical assessment is based on retinal examination and digital fundus imaging. When determining the severity of a disease and the state of therapy, ophthalmologists take into account characteristics such vascular dilation, tortuosity, demarcation lines, and ridges. Practical constraints can have an impact on the screening procedure. Image quality can be lowered by infant movement and media opacity, and grading judgments can be impacted by observer differences . Another limitation is access to expert evaluation, especially when a significant number of preterm babies need to have repeated exams. Clinical environments also differ in practice patterns, such as how ROP is treated and monitored . Automated retinal image analysis has been investigated as a way to assist image interpretation and screening. CNN-based methods can learn discriminative retinal features directly from labeled images, and related work spans diabetic-retinopathy analysis as well as ROP-specific applications . The ROP literature also includes studies of associated ophthalmic complications and the training required for reliable clinical assessment . These strands of work reflect different clinical needs, from image-based classification to specialist interpretation and follow-up. In the present study, automated classification is considered as decision support rather than a replacement for ophthalmic assessment. Transfer learning provides a practical way to adapt deep neural networks when task-specific medical image datasets are smaller than the datasets typically used to train large models from random initialization. A network pretrained on a large image collection such as ImageNet can be adapted to a target task by replacing or modifying its classification layers and fine-tuning part of the network. Transfer learning has been applied to medical classification problems outside ROP, including MRI-based disease classification and diabetes detection . Deep-learning methods have also been studied specifically for ROP screening . In the present work, pretrained networks are fine-tuned on the same ROP dataset. The models are then compared using the same data partitions, preprocessing steps, training procedure, and evaluation measures. Previous reviews of AI methods for ROP report differences in the prediction tasks, datasets, network architectures, and validation procedures used across papers . Both image classification and retinal blood-vessel analysis have been used in ROP research . Papers also vary in the quantity and description of output categories. While the current task distinguishes between Healthy, Type 1 ROP, Type 2 ROP, and RD, other models employ fewer diagnostic classifications. It is challenging to directly compare results from different research due to differences in image sources and dataset makeup. Different clinical concerns are addressed by work on other retinal disorders, which should be separated from ROP-specific categorization research . Because of this, the preprocessing, augmentation, data partitioning, and evaluation criteria used in our trials are the same for all six models. Four groups are included in the categorization task: RD, Type 1 ROP, Type 2 ROP, and Healthy. Using the same experimental setup, we evaluate AlexNet, MobileNetV2, VGG19, DenseNet121, EfficientNetV2-S, and CBAM-ResNet50. The CBAM-ResNet50 configuration incorporates the Convolutional Block Attention Module (CBAM), which applies channel and spatial attention to the feature maps produced by the underlying ResNet50 architecture. Image preprocessing and augmentation are applied before training to account for variation in illumination, retinal appearance, and image acquisition conditions.

Figure 1 contrasts the main stages of conventional ROP screening with the image-processing process considered in this paper. Conventional screening depends on specialist examination and manual interpretation of retinal findings. In the proposed workflow, preprocessing, feature extraction, and classification are performed computationally, after which the predicted class can be used as decision-support information. The final clinical interpretation remains with the ophthalmologist.

The main contributions of this paper are as follows:

  • We formulate a transfer-learning framework for four-class ROP classification using pretrained CNN architectures and retinal fundus images.

  • We evaluate a CBAM-ResNet50 configuration in which channel and spatial attention are incorporated into ResNet50 to emphasize retinal features used during classification.

  • We apply a preprocessing and data-augmentation process designed for neonatal retinal images, including variation associated with illumination, motion artifacts, and retinal appearance.

  • We compare six pretrained architectures, namely AlexNet, MobileNetV2, VGG19, DenseNet121, EfficientNetV2-S, and CBAM-ResNet50, using the same four-class dataset and 5-fold cross-validation. Each model was tested on five held-out partitions rather than on one fixed train-test split. The same partitioning and training procedure was used for all six architectures.

  • Classification performance is reported using accuracy, sensitivity, specificity, precision, F1-score, AUC, confusion matrices, and ROC curves. The class-wise measures and confusion matrices show where errors occur, while AUC and ROC curves describe discrimination across decision thresholds.

The rest of the paper is arranged as follows. Section 2 discusses previous work on ROP detection, classification, and related prediction tasks. Section 3 gives the dataset, preprocessing and augmentation steps, model architectures, training procedure, and evaluation protocol. The experimental results and their analysis are presented in Section 4. Section 5 closes the paper with the main findings, limitations, and directions for further work.

Complete Article

The complete article, including all figures, tables, equations and algorithms, is available in the official publication PDF.

Conclusion

We compared six pretrained CNN configurations for four-class ROP classification from retinal fundus images. The four classes were Healthy, Type 1, Type 2, and RD. Each model was trained with transfer learning and evaluated using 5-fold cross-validation with training-time image augmentation. CBAM-ResNet50 recorded the highest values in Table 3, with 95.4% accuracy, 96.1% sensitivity, 94.8% specificity, 95.7% precision, and a 95.9% F1-score. EfficientNetV2-S achieved the second-highest accuracy at 93.2%. The confusion-matrix results also show fewer classification errors for CBAM-ResNet50 in the displayed evaluation than for the other models. These comparisons indicate that CBAM-ResNet50 performed best under the experimental conditions used in this paper. However, the present experiments do not isolate the contribution of the attention module from other architectural differences. The fold-specific ROC analysis also shows that performance was not uniform across data partitions. CBAM-ResNet50 reached a macro-AUC of 0.975 in Fold 5 and 0.942 in Fold 4, while the lowest reported value was 0.700 in Fold 1. The mean macro-AUC across the five folds was 0.854. This variation indicates that performance remains sensitive to the composition of the evaluation fold and should therefore be examined on additional external clinical datasets before conclusions about generalisation are made. The current results support further investigation of transfer-learning-based ROP classification as a decision-support approach. Clinical deployment, real-time operation, and reduction of screening workload were not evaluated directly in this paper and therefore remain subjects for future validation.

Future work will examine longitudinal retinal image sequences for disease-progression modelling and early-stage risk prediction. Additional clinical datasets with greater variation in patient characteristics and acquisition conditions will be considered to assess generalisation across sites. Domain adaptation will also be investigated to address differences between imaging environments. Finally, model compression and quantisation will be studied for deployment on mobile and edge devices and for possible integration with teleophthalmology systems.

References

  1. O. Dammann, M. E. Hartnett, and A. Stahl, “Retinopathy of prematurity, ” Developmental Medicine and Child Neurology, vol. 65, no. 5, pp. 625-631, 2023, doi: 10.1111/dmcn.15468
  2. M. W. Nadeem, H. G. Goh, M. Hussain, S.-Y. Liew, I. Andonovic, and M. A. Khan, “Deep learning for diabetic retinopathy analysis: A review, research challenges, and future directions, ” Sensors, vol. 22, no. 18, p. 6780, 2022, doi: 10.3390/s22186780
  3. E. Koc and A. Y. Bas, “Retinopathy of prematurity (ROP): From the perspective of the neonatologist, ” Global Pediatrics, vol. 8, p. 100159, 2024, doi: 10.1016/j.gpeds.2024.100159
  4. N. Shaikh et al., “Glaucoma in retinopathy of prematurity: A review, ” Survey of Ophthalmology, 2025, doi: 10.1016/j.survophthal.2025.03.006
  5. V. M. R. Sankari, U. Snekhalatha, A. Chandrasekaran, and P. Baskaran, “Automated diagnosis of retinopathy of prematurity from retinal images of preterm infants using hybrid deep learning techniques, ” Biomedical Signal Processing and Control, vol. 85, p. 104883, 2023, doi: 10.1016/j.bspc.2023.104883
  6. A. Song et al., “Retinopathy of prematurity training and education: A systematic review, ” Survey of Ophthalmology, 2025, doi: 10.1016/j.survophthal.2025.06.007
  7. S. S. Hussain, P. M. Shah, H. Dawood, and et al., “A swin transformer and CNN fusion framework for accurate parkinson disease classification in MRI, ” Scientific Reports, vol. 15, p. 15117, 2025, doi: 10.1038/s41598-025-93671-5
  8. F. Rustam, A. S. Al-Shamayleh, R. Shafique, and et al., “Enhanced detection of diabetes mellitus using novel ensemble feature engineering approach and machine learning model, ” Scientific Reports, vol. 14, p. 23274, 2024, doi: 10.1038/s41598-024-74357-w
  9. V. Kumar, H. Patel, K. Paul, and S. Azad, “Deep learning-assisted retinopathy of prematurity (ROP) screening, ” ACM Transactions on Computing for Healthcare, vol. 4, 2023, doi: 10.1145/3596223
  10. R. Shemesh, M. Chiang, R. V. P. Chan, and et. al, “An international survey on retinopathy of prematurity practice patterns during the COVID-19 pandemic and lessons for future management, ” International Ophthalmology, vol. 44, p. 407, 2024, doi: 10.1007/s10792-024-03290-8
  11. R. H. Gensure, M. F. Chiang, and J. P. Campbell, “Artificial intelligence for retinopathy of prematurity, ” Current Opinion in Ophthalmology, vol. 31, no. 5, pp. 312-317, Sep. 2020, doi: 10.1097/ICU.0000000000000680
  12. A. Jafarizadeh, S. F. Maleki, P. Pouya, and et. al, “Current and future roles of artificial intelligence in retinopathy of prematurity, ” Artificial Intelligence Review, vol. 58, p. 188, 2025, doi: 10.1007/s10462-025-11153-6
  13. A. D'Angelo, F. Lixi, L. Vitiello, V. Gagliardi, A. Pellegrino, and G. Giannaccare, “The role of diet and oral supplementation for the management of diabetic retinopathy and diabetic macular edema: A narrative review, ” BioMed Research International, vol. 2025, p. 6654976, 2025, doi: 10.1155/bmri/6654976
  14. M. K, S. M. John, A. Joseph, and B. Abraham, “A study on retinopathy of prematurity(ROP) screening using deep learning approaches and a blood vessel segmentation framework for ROP affected eyes, ” in 2024 second international conference on emerging trends in information technology and engineering (ICETITE), 2024, pp. 1-10. doi: 10.1109/ic-ETITE58242.2024.10493798
  15. A. H. Khan, H. Malik, W. Khalil, S. K. Hussain, T. Anees, and M. Hussain, “Spatial correlation module for classification of multi-label ocular diseases using color fundus images, ” Computers, Materials and Continua, vol. 76, no. 1, pp. 133-150, 2023.
  16. J. Chen, Y. Zhu, L. Li, and et. al, “Visual impairment burden in retinopathy of prematurity: Trends, inequalities, and improvement gaps, ” European Journal of Pediatrics, vol. 183, pp. 1891-1900, 2024, doi: 10.1007/s00431-024-05450-5
  17. H. Hashemian et al., “Application of artificial intelligence in ophthalmology: An updated comprehensive review, ” Journal of Ophthalmic and Vision Research, vol. 19, pp. 354-367, 2024, doi: 10.18502/jovr.v19i3.15893
  18. S. Mushriff and et. al, “Risk factors associated with retinopathy of prematurity, ” International Journal of Health Sciences, vol. 6, no. S9, pp. 4895-4909, 2022, doi: 10.53730/ijhs.v6nS9.14198
  19. Y. Zhong, X. Lin, Y. Yang, and et. al, “Screening for retinopathy of prematurity in china: A five-year cohort study in seven screening centers, ” BMC Ophthalmology, vol. 25, p. 3, 2025, doi: 10.1186/s12886-024-03836-5
  20. A. T. Wang and S. Dai, “Preferred treatment patterns of retinopathy of prematurity: An international survey, ” Pediatric Reports, vol. 16, no. 3, pp. 816-822, 2024, doi: 10.3390/pediatric16030069
  21. H. Sun et al., “Identification of genetic factors underlying severe retinopathy of prematurity in preterm infants, ” Molecular Vision, vol. 31, pp. 33-43, 2025.
  22. P. Huang et al., “ROPRNet: Deep learning-assisted recurrence prediction for retinopathy of prematurity, ” Biomedical Signal Processing and Control, vol. 100, p. 107135, 2025, doi: 10.1016/j.bspc.2024.107135
  23. W. Yang et al., “An interpretable system for screening the severity level of retinopathy in premature infants using deep learning, ” Bioengineering, vol. 11, no. 8, p. 792, 2024, doi: 10.3390/bioengineering11080792
  24. B. Ebrahimi et al., “Assessing spectral effectiveness in color fundus photography for deep learning classification of retinopathy of prematurity, ” Journal of Biomedical Optics, vol. 29, no. 7, p. 076001, 2024, doi: 10.1117/1.JBO.29.7.076001
  25. X. Luo, S. Zhao, Y. Chen, G.-S. Ying, and L. He, “VGG-ST: A VGG-swin transformer-based model for ROP disease diagnosis, ” in 2024 IEEE international conference on bioinformatics and biomedicine (BIBM), 2024, pp. 4987-4994. doi: 10.1109/BIBM62325.2024.10822551
  26. V. M. R. Sankari, U. Umapathy, S. Alasmari, and S. M. Aslam, “Automated detection of retinopathy of prematurity using quantum machine learning and deep learning techniques, ” IEEE Access, vol. 11, pp. 94306-94321, 2023, doi: 10.1109/ACCESS.2023.3311346
  27. Y. Peng et al., “Automatic zoning for retinopathy of prematurity with a key area location system, ” Biomedical Optics Express, vol. 15, no. 2, pp. 725-742, 2024, doi: 10.1364/BOE.506119
  28. N. Salih, M. Ksantini, N. Hussein, D. Ben Halima, A. Abdul Razzaq, and S. Ahmed, “Deep learning models and fusion classification technique for accurate diagnosis of retinopathy of prematurity in preterm newborn, ” Baghdad Science Journal, vol. 21, no. 5, 2024, doi: 10.21123/bsj.2023.8747
  29. M. Yurdakul, K. Uyar, S. Tasdemir, and I. Atabas, “ROPGCViT: A Novel Explainable Vision Transformer for Retinopathy of Prematurity Diagnosis, ” IEEE Access, vol. 13, pp. 77064-77079, 2025, doi: 10.1109/ACCESS.2025.3564213
  30. Y. Chu et al., “Image analysis-based machine learning for the diagnosis of retinopathy of prematurity: A meta-analysis and systematic review, ” Ophthalmology Retina, vol. 8, no. 7, pp. 678-687, 2024, doi: 10.1016/j.oret.2024.01.013
  31. C. Durmaz Engin, T. Ozturk, O. Ozkan, and et al., “Prediction of retinopathy of prematurity development and treatment need with machine learning models, ” BMC Ophthalmology, vol. 25, p. 194, 2025, doi: 10.1186/s12886-025-04025-8
  32. B. Young et al., “Deep learning-based prediction of genetic risk in retinopathy of prematurity using retinal images, ” Investigative Ophthalmology & Visual Science, vol. 65, no. 7, p. 3764, 2024.
  33. A. S. Govind, A. Gadad, D. J. Garodia, A. Gupta, A. Vinekar, and G. Srinivasa, “Deep learning for classifying stages of retinopathy of prematurity, ” in 2024 international conference on advances in modern age technologies for health and engineering science (AMATHE), Shivamogga, India, 2024, pp. 1-6. doi: 10.1109/AMATHE61652.2024.10582102
  34. S. Chen, X. Zhao, Z. Wu, and et. al, “Multi-risk factors joint prediction model for risk prediction of retinopathy of prematurity, ” EPMA Journal, vol. 15, pp. 261-274, 2024, doi: 10.1007/s13167-024-00363-7
  35. G. Hubert and S. Silvia Priscila, “Comprehensive prediction of retinopathy in preterm infants using deep learning approaches, ” in Advancing intelligent networks through distributed optimization, S. S. Rajest and others, Eds., Hershey, PA: IGI Global Scientific Publishing, 2024, pp. 353-370. doi: 10.4018/979-8-3693-3739-4.ch018
  36. A. Ortiz, S. Patino, J. Torres, et al., “AI-enabled screening for retinopathy of prematurity in low-resource settings, ” JAMA Network Open, vol. 8, no. 4, p. e257831, 2025, doi: 10.1001/jamanetworkopen.2025.7831
  37. D. Lakshmanan, N. R. S. Raaj, S. Ravi, M. Ayyaswamy, and S. Ragunathan, “Retinopathy detection of premature using machine learning algorithms, ” in AIP conference proceedings, AIP Publishing, 2025, p. 030001. doi: 10.1063/5.0271997
  38. J. Timkovic, J. Nowakova, J. Kubicek, and et. al, “Retinal image dataset of infants and retinopathy of prematurity, ” Scientific Data, vol. 11, p. 814, 2024, doi: 10.1038/s41597-024-03409-7
  39. A. S. Coyner et al., “Use of an artificial intelligence-generated vascular severity score improved plus disease diagnosis in retinopathy of prematurity, ” Ophthalmology, vol. 131, no. 11, pp. 1290-1296, 2024, doi: 10.1016/j.ophtha.2024.06.006
  40. K. Deepthi, M. S. Josephine, and V. Jayabala Raja, “Automated diagnosis of plus disease in retinopathy of prematurity based on transformer-based unsupervised curriculum learning, ” Biomedical Signal Processing and Control, vol. 104, p. 107521, 2025, doi: 10.1016/j.bspc.2025.107521