Acquire continuous and precise score for fundus image quality assessment: FTHNet and FQS dataset.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal fundus images are crucial for medical diagnosis, but their quality significantly impacts accuracy.
• Existing fundus image quality assessment (FIQA) methods lack the detailed granularity needed for clinical applications due to dataset and algorithmic limitations.

Purpose of the Study:

• To introduce a comprehensive benchmark dataset for FIQA, named Fundus Quality Score (FQS).
• To develop a novel Transformer-based Hypernetwork (FTHNet) for precise FIQA, treating it as a regression task.

Main Methods:

• Created the FQS dataset with 2,246 images, annotated with continuous mean opinion scores (0-100) and quality categories.
• Designed FTHNet, a Transformer-based Hypernetwork, to predict continuous quality scores (MOS) rather than using classification.

Main Results:

• FTHNet achieved high performance on the FQS dataset, with Pearson (0.9423) and Spearman (0.9488) correlation coefficients.
• The proposed method significantly outperformed existing approaches with reduced parameters and computational cost.
• Model deployment indicated potential for automated medical image quality control.

Conclusions:

• The FQS dataset and FTHNet offer a significant advancement in fundus image quality assessment.
• FTHNet provides a more granular and accurate method for FIQA, suitable for clinical workflows.
• The release of code and data will foster further research in automated medical image analysis.