Abstract
Retinopathy of Prematurity (ROP) is the major cause of blindness in preterm infants, making early detection and precise classification essential for effective intervention. This research analyzes the effectiveness of four transfer learning models (InceptionV3, DenseNet201, MobileNet, and Xception), along with several voting-based ensemble approaches, such as Soft Voting, Hard Voting, Weighted Average Voting, and Mix Voting for identifying ROP in binary and multi-class conditions. A novel dataset obtained from Aravind Eye Hospital, Chennai, was utilized for this classification task. Among the individual models, Xception performed better, with an F1 score of 96.95% and an accuracy of 97.28%. To improve prediction performance, base models were combined using ensemble techniques. Soft Voting and Mix Voting achieved an accuracy of 98.37%, an F1 score of 98.14%, and a specificity of 99.02% in binary classification. These findings suggest that ensemble techniques can greatly improve the accuracy and reliability of ROP classification, thereby facilitating improved clinical decision-making.
The Challenge of ROP Screening
Retinopathy of Prematurity (ROP) occurs when abnormal blood vessels grow in the retina of preterm infants, potentially leading to retinal detachment and permanent blindness. While early detection can prevent vision loss, there is a massive "expert gap." In India, for instance, over 65% of the population lives in rural areas, yet fewer than 100 ROP experts are available nationwide to perform screenings.
The challenge lies in creating an automated diagnostic tool that is both highly accurate—to avoid missing cases—and reliable enough for clinical decision-making. Existing single-model AI approaches often struggle with the subtle features of early-stage ROP or produce conflicting results. This research seeks to bridge that gap using ensemble learning to improve diagnostic consistency.
Proposed Methodology
The study employed a multi-stage pipeline involving Transfer Learning and Ensemble Voting Classifiers. The researchers fine-tuned four established convolutional neural network (CNN) architectures: InceptionV3, DenseNet201, MobileNet, and Xception. These models were chosen for their diverse strengths, ranging from MobileNet’s efficiency to Xception’s depthwise separable convolutions.
To combine the predictions of these four models, the authors compared several voting mechanisms:
- Hard Voting: Majority rule based on class labels.
- Soft Voting: Averaging the predicted probabilities across models.
- Weighted Average Voting (WAV): Assigning more "weight" to models with higher individual performance.
- Mix Voting: A novel hybrid approach designed to resolve "tie" scenarios in Hard Voting (e.g., when two models say "ROP" and two say "No ROP") by deferring to Soft Voting probabilities to break the deadlock.
Dataset and Feature Identification
The study utilized a novel dataset of 610 high-quality retinal fundus images obtained from the Aravind Eye Hospital in Chennai, India. The images were categorized for both binary classification (ROP vs. No ROP) and multi-class classification (Advanced ROP, Early ROP, and No ROP).
To ensure the AI was making decisions based on medically relevant features, the researchers used LIME (Local Interpretable Model-agnostic Explanations). This visualization confirmed that the models focused on critical areas like the "ridge" or demarcation line—the boundary between vascularized and non-vascularized retina.
Key Results and Clinical Implications
The ensemble approach showed marked improvements over individual models. In binary classification, both Soft Voting and Mix Voting achieved a peak accuracy of 98.37% and an F1 score of 98.14%. Crucially, the ensemble reached a specificity of 99.02%, effectively minimizing false alarms that lead to unnecessary parental anxiety and hospital costs.
In the more difficult task of multi-class classification, the ensemble models maintained high performance, particularly in identifying "Advanced ROP" with near-perfect precision. Statistical analysis (ANOVA and Tukey’s HSD tests) confirmed that the performance gains of the ensemble methods over individual models were statistically significant ($p < 0.001$).
Conclusion and Future Directions
By combining multiple transfer learning models through a novel "Mix Voting" strategy, this framework provides a robust solution for early ROP detection. While the results are promising, the authors plan to expand the dataset to include more diverse images and incorporate Stage 4 and Stage 5 ROP cases, which involve partial or total retinal detachment. Future iterations may also explore "stacking" approaches, where a second-level model learns how to best combine the outputs of the base CNNs.
Summary updated