Joint retina segmentation and classification for early glaucoma diagnosis.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma diagnosis relies on analyzing the retinal nerve fiber layer (RNFL) from optical coherence tomography (OCT) images.
• Current diagnostic methods may benefit from automated analysis to improve accuracy and efficiency.

Purpose of the Study:

• To develop and validate a joint deep learning model for simultaneous retinal layer segmentation and glaucoma classification using OCT images.
• To simulate the clinical decision-making process of ophthalmologists in diagnosing glaucoma.

Main Methods:

• A novel deep model comprising a segmentation network and a classification network was proposed.
• The segmentation network predicts six retinal layers and five boundaries, followed by a topology-correcting post-processing algorithm.
• The classification network utilizes the RNFL thickness vector to determine glaucoma probability, incorporating a clinically inspired diagnostic module.

Main Results:

• The proposed method demonstrated superior retinal layer segmentation performance compared to state-of-the-art techniques on both collected and public datasets.
• The glaucoma classification achieved a diagnostic accuracy of 81.4% with an Area Under the Curve (AUC) of 0.864.
• The model outperformed baseline methods in glaucoma classification accuracy.

Conclusions:

• The joint segmentation and classification deep model effectively aids in early glaucoma diagnosis using OCT imaging.
• The model's ability to accurately segment retinal layers and classify glaucoma shows significant promise for clinical application.
• This approach offers a potential advancement in automated glaucoma detection and management.