Deep learning-based automatic image quality assessment in ultra-widefield fundus photographs

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Ultra-widefield (UWF) fundus photography is increasingly vital for clinical diagnosis and AI development.
• Accurate image quality assessment (IQA) is a critical prerequisite for reliable UWF image interpretation.
• Current IQA methods rely on manual grading, which can be subjective and time-consuming.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for automated IQA of UWF fundus photographs (UWF-IQA model).
• To compare the performance of the automated UWF-IQA model against manual grading by ophthalmologists.
• To assess intergrader agreement among human experts in UWF image quality assessment.

Main Methods:

• A dataset of 4749 UWF images was curated.
• Three ophthalmologists manually graded images based on field of view, peripheral visualization, posterior pole details, and centring using a five-point scale.
• An EfficientNet-B3-based DL model (UWF-IQA model) was developed to predict IQA scores.

Main Results:

• The UWF-IQA model achieved an average agreement of 0.731 (Cohen's quadratic weighted kappa) with manual graders.
• Manual graders achieved an average intergrader agreement of 0.603.
• The model showed highest agreement in posterior pole grading (0.838) and final grading (0.788).

Conclusions:

• The automated UWF-IQA model demonstrates superior agreement with manual assessments compared to intergrader agreement.
• The DL-based UWF-IQA model provides robust and efficient predictions for both final and subcategory gradings.
• This automated approach can enhance the consistency and efficiency of IQA in clinical settings and AI applications.