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

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Artificial Intelligence in Healthcare

Background:

• Multilabel classification of retinal diseases is crucial for early diagnosis.
• Lightweight convolutional neural networks (CNNs) show promise but require careful architectural consideration.
• Evaluating CNNs beyond predictive performance is essential for clinical translation.

Purpose of the Study:

• To investigate the influence of architectural design on lightweight CNN behavior in retinal disease classification.
• To analyze multiple performance dimensions beyond accuracy, including stability, agreement, and error complementarity.
• To explore the potential of architectural diversity for developing robust diagnostic systems.

Main Methods:

• Utilized the ODIR-5K dataset for multilabel classification of retinal diseases.
• Assessed lightweight CNNs across dimensions: accuracy, training stability, inter-model agreement, error complementarity, and image-level consensus.
• Analyzed prediction patterns and failure modes across different architectures.

Main Results:

• No single CNN architecture achieved optimal performance and stability simultaneously.
• Architectures exhibited distinct prediction patterns and complementary failure modes on disjoint data subsets.
• Model disagreement correlated with classification difficulty, indicating potential as an uncertainty proxy.
• Architectural inductive bias significantly shapes CNN behavior beyond aggregate metrics.

Conclusions:

• Architectural diversity in lightweight CNNs is key for improving performance and reliability.
• Model combination strategies, leveraging architectural differences, can enhance diagnostic decision support.
• Understanding diverse model behaviors is critical for developing robust AI in medical imaging.