Advancing Glaucoma Detection With Ai-Driven Analysis Of Color Fundus Photographs

Published 2024 - 42nd Congress of the ESCRS

Reference: PP09.11 | Type: Free paper | DOI: 10.82333/f936-t212

Authors: Volga Likhachevskaya* ¹ , Ina Malinouskaya ² , Anna Furmanczuk ³ , Alexei Kuzmenkov ¹

¹Deepdee BV,The Hague,Netherlands, ²ophthalmological,BSMU,Minsk,Belarus, ³BCM klinika okulistyki,Bielsko-Biała,Poland

Purpose

To enhance artificial intelligence (AI) algorithms for glaucoma detection, particularly emphasizing multi-label classification of glaucomatous features in color fundus photographs (CFPs), crucial for timely diagnosis and management to mitigate glaucoma-related vision loss. This research aims to improve the accuracy and efficiency of glaucoma screening methods, ultimately leading to earlier diagnosis and appropriate treatment, thereby minimizing the risk of irreversible vision impairment associated with the disease.

Setting

Using a dataset of more than 110,000 annotated CFPs, the study aims to improve AI algorithms to identify ten critical glaucomatous features that promote early diagnosis and effective treatment to reduce the prevalence of glaucoma-related blindness and loss of visual function. Testing AI models, allowing comprehensive analyses and validation of the performance of the developed algorithms in different clinical scenarios and on different patient cohorts.

Methods

The dataset comprises 101,442 training images and 9,741 testing images categorized as "referable glaucoma" (RG) or "no referable glaucoma" (NRG). RG images are annotated with specific glaucomatous features. AI algorithms are developed for binary (RG vs. NRG) and multi-label classification to identify glaucomatous features.

Binary Classification Task: Efficiently segregating fundus images into RG and NRG categories streamlines referrals for at-risk patients.

Multi-Label Classification Task: Identifying ten specific glaucomatous features enhances diagnostic accuracy and understanding of the disease.

Results

The Deepdee algorithms achieved high accuracy in binary classification: Accuracy: 95%, Specificity: 94%, Sensitivity: 96%. For multi-label classification, Macro-Averaged Precision: is 85%, and Macro-Averaged Recall: is 80%, indicating precise identification of glaucomatous features. These results demonstrate the effectiveness and reliability of the AI-driven approach in detecting glaucoma and its associated features, highlighting its potential as a valuable tool in clinical practice for early diagnosis and appropriate management.

Conclusions

This study underscores AI's potential to revolutionize glaucoma screening, ensuring early diagnosis and management, and ultimately reducing glaucoma-related blindness and vision function loss. Integrating AI into screening programs promises advanced diagnostic tools, improving patient outcomes. Discussions on metrics like accuracy, precision, recall, and AUC evaluate algorithm effectiveness. AI shows promise in early glaucoma detection, offering hope for reducing blindness incidence. Further research directions will focus on the integration of Deepdee AI solutions into clinical practice, assessing their impact on patient care and outcomes.