MAPLES-DR: MESSIDOR Anatomical and Pathological Labels for Explainable Screening of Diabetic Retinopathy.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) and Macular Edema (ME) require early diagnosis to prevent vision loss.
• Explainability of AI screening models for DR and ME remains a significant research challenge.
• Existing datasets may not fully support the development of explainable AI for these conditions.

Purpose of the Study:

• To introduce MAPLES-DR, a novel dataset designed to facilitate research in explainable AI for DR and ME screening.
• To provide detailed annotations, including pixel-wise segmentation maps of 10 DR-related biomarkers, for a subset of the MESSIDOR dataset.
• To document the dataset creation process, assess annotation quality, and offer usage guidelines for machine learning applications.

Main Methods:

• Developed MAPLES-DR by adding new DR and ME diagnoses and pixel-wise segmentation maps to 198 images from the public MESSIDOR dataset.
• Focused on 10 anatomical and pathological biomarkers crucial for DR assessment.
• Documented annotation procedures, evaluated interobserver variability, and ensured annotation quality.

Main Results:

• Created MAPLES-DR, a resource with enhanced diagnostic and segmentation information for DR and ME research.
• The dataset includes detailed annotations for 198 fundus images, supporting explainable AI development.
• Analysis confirmed the quality and reliability of the annotations, with discussed interobserver variability.

Conclusions:

• MAPLES-DR provides a valuable resource for advancing explainable AI in the automated diagnosis of Diabetic Retinopathy and Macular Edema.
• The dataset's detailed annotations and segmentation maps will aid in developing more transparent and reliable screening tools.
• This work contributes to improving patient monitoring and enabling earlier interventions for DR and ME, ultimately preserving vision.