AngioReport: dataset and baseline methods for fundus angiography report generation

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Fundus angiography, including fundus fluorescein angiography (FFA) and indocyanine green angiography (ICGA), is crucial for diagnosing retinal diseases.
• Developing automated systems for analyzing these complex images can improve diagnostic efficiency and accuracy.

Purpose of the Study:

• To create the largest annotated fundus angiographic dataset to date.
• To establish baseline artificial intelligence (AI) methods for automatic report generation from FFA and ICGA images.
• To facilitate advancements in AI applications within ophthalmology.

Main Methods:

• A retrospective review of 55,361 de-identified fundus angiographic images from 1691 patients (≥18 years) was conducted.
• Retinal specialists annotated images detailing 24 diagnostic conditions, including type, location, and fluorescence patterns.
• Two AI models were developed: a classification-based approach (ResNet101) for lesion recognition and a language-generation approach (Bootstrapping Language-Image Pre-training) for report generation.

Main Results:

• The dataset comprises 24 diagnostic conditions, with macular neovascularisation being most common.
• The classification-based AI model achieved an F1 score of 7.966, excelling in recognizing specific lesions and fluorescence.
• The language-generation AI model achieved a BERTScore of 7.947, demonstrating strength in impression recognition and hyperfluorescence identification.

Conclusions:

• The developed dataset is the largest annotated fundus angiographic dataset available.
• Both baseline AI methods show significant promise for automated report generation in ophthalmology.
• This work provides a strong foundation for future AI-driven diagnostic tools in retinal imaging.