Architectural inductive bias in from-scratch CNN training for multi-disease fundus image classification.

This study investigates how architectural design influences the behavior of lightweight convolutional neural networks (CNNs) in multilabel classification of retinal diseases using the ODIR-5K dataset. Rather than focusing solely on predictive performance, we analyze multiple complementary dimensions, including accuracy, stability across training runs, inter-model agreement, error complementarity, and image-level consensus. Results show that no single architecture consistently achieves both optimal performance and stability. Moreover, different architectures exhibit distinct prediction patterns and fail on partially disjoint subsets of the data, leading to significant complementarity. Agreement and consensus analyses further reveal that disagreement across models correlates with classification difficulty, suggesting its potential as a proxy for uncertainty. These findings demonstrate that architectural inductive bias plays a central role in shaping model behavior beyond aggregate metrics. Importantly, the observed diversity enables improved performance and reliability through model combination, supporting the development of ensemble-based systems for clinically relevant decision support.