Optimizing Field-of-View for Type-1 Retinopathy of Prematurity via Multiple Instance Learning

Purpose: To cut down on screening time and costs, this study optimizes field-of-view (FOV) combinations for classifying type-1 retinopathy of prematurity (ROP) through multiple-instance learning that models clinical decision-making. Methods: The dataset included 204 eyes from Chang Gung Memorial Hospital (CGMH) and 80 eyes from Osaka University, each eye represented by five FOVs (temporal, nasal, central, superior, inferior). To optimize type-1 ROP classification, we employed multiple instance learning (MIL) at the eye-level, integrating information from a subset of distinct retinal FOV. A CNN-based feature extractor was used to acquire features for image representation. We systematically evaluated various FOV combinations and compared two fusion strategies—feature-level and outcome-level—to identify the most effective approach. Results: Model performance was evaluated using five-fold cross-validation, with metrics including accuracy, precision, recall and F1-score. A t-test was used to analyze the statistical differences between outcomes. The feature-level MIL significantly outperformed the outcome-level MIL in all combinations and metrics ( p < 0.05) except for vertical views. Among FOV combinations, the temporal, nasal and central set achieved the highest performance (accuracy: 0.888, F1-score: 0.825). Horizontal (temporal and nasal) and central views individually showed strong diagnostic power, comparable to multi-view settings, whereas vertical (superior and inferior) views performed significantly worse. Conclusions: The feature-level Multiple Instance Learning (MIL) model effectively emulates clinical decision-making for type-1 ROP classification. By concentrating on temporal, nasal, and central views, it achieved superior performance, significantly reducing image counts without compromising accuracy. Notably, feature-level MIL outperformed outcome-level MIL, underscoring the critical role of information fusion strategy. This approach promises to enhance the practicality and efficiency of AI-assisted ROP diagnosis, offering a tangible improvement for patient care in telemedicine.