Retinal image quality assessment using deep learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Poor-quality retinal images hinder accurate medical diagnosis.
• Retaking fundus photography exams is inconvenient for patients.
• Automated quality assessment can assist healthcare professionals during image acquisition.

Purpose of the Study:

• To propose a robust automatic system for assessing retinal image quality at the point of acquisition.
• To aid healthcare professionals in ensuring high-quality fundus photography.
• To develop a reliable tool for real-time image quality evaluation.

Main Methods:

• A convolutional neural network (CNN) was pretrained on non-medical images for feature extraction.
• The CNN underwent fine-tuning with a small dataset of labeled retinal images.
• Performance was validated on DRIMDB and ELSA-Brasil databases using cross-validation.

Main Results:

• The CNN achieved an Area Under the Curve (AUC) of 99.98% on the DRIMDB database.
• The CNN achieved an AUC of 98.56% on the ELSA-Brasil database in inter-database testing.
• The model demonstrated robustness across different image acquisition scenarios.

Conclusions:

• The proposed CNN-based system effectively assesses retinal image quality in real-time.
• The system's robustness makes it suitable for clinical deployment.
• This technology can enhance diagnostic accuracy and streamline ophthalmic examinations.