A comparative study of supervised, transfer learning, and anomaly detection methods for fundus image quality assessment

• Autores: Telesco, Lucas Gabriel; Peralta, César Sebastián
• Año de publicación: 2025
• Idioma: inglés
• Tipo de recurso: documento de conferencia
• Materia: Ciencias Informáticas; Fundus image quality assessment; Deep learning; OneClass classification; Supervised learning; Transfer learning
• Institución: Universidad Nacional de La Plata

Retinal image quality assessment (RIQA) is a fundamental step for manual or automated diagnosis, but supervised methods require large volumes of labeled data. This paper presents a comparative study between a supervised approach, transfer learning-based methods, and anomaly detection for the RIQA task. A supervised ResNet-18 architecture was evaluated against one-class classification models (OneClass SVM and Deep SVDD) on the public DDR and Kaggle (EyePACS) datasets. The results show that the supervised model achieves the highest performance in in-domain evaluations (F1-Score of 0.914 on DDR), but suffers a significant degradation in unseen domains. The transfer learning-based approach (ResNet+SVM) proved to be the most effective unsupervised strategy (F1-Score of 0.670) and a computationally efficient starting point.
Red de Universidades con Carreras en Informática