Wen-Hsi Lan, Chia-Ling Tsai, Wei-Chi Wu
The feature-level MIL model effectively simulates clinical decision making for type-1 ROP classification. Temporal, nasal, and central views, as well as their combination, achieved superior performance over vertical views or their combination. Notably, feature-level MIL outperformed outcome-level MIL, underscoring the critical role of information fusion strategy.
PURPOSE: The purpose of this study was 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 (MIL) that models clinical decision making.
METHODS: The dataset included 1420 photographs of 284 eyes (204 eyes from Chang Gung Memorial Hospital [CGMH] and 80 eyes from Osaka University), each eye containing five FOVs (temporal, nasal, central, superior, and inferior). We evaluated various FOV combinations and compared two MIL fusion strategies, feature-level and outcome-level, to identify the most effective approach. Feature-level learns a joint representation across views, whereas outcome-level aggregates the predicted probabilities from each view at the decision stage. Model performance was evaluated using five-fold cross-validation, with metrics including accuracy, precision, recall, F1-score, and area under the curve (AUC).
RESULTS: The feature-level MIL significantly outperformed the outcome-level MIL in all combinations and metrics except the superior view (P < 0.05). Among FOV combinations, the temporal, nasal, and central set achieved the highest performance (accuracy = 0.892 ± 0.043, F1-score = 0.831 ± 0.097). Horizontal (temporal and nasal) and central views individually showed stronger diagnostic power, comparable to multi-view settings, whereas vertical (superior and inferior) views performed significantly worse.
CONCLUSIONS: The feature-level MIL model effectively simulates clinical decision making for type-1 ROP classification. Temporal, nasal, and central views, as well as their combination, achieved superior performance over vertical views or their combination. Notably, feature-level MIL outperformed outcome-level MIL, underscoring the critical role of information fusion strategy.
TRANSLATIONAL RELEVANCE: Clinically aligned MIL with optimized FOV selection enables more efficient artificial intelligence (AI)-assisted telemedicine screening for retinopathy of prematurity.