Rim learning framework based on TS-GAN: A new paradigm of automated glaucoma screening from fundus images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma detection often uses Cup-to-Disc Ratio (CDR) and Rim-to-Disc Ratio (RDR), but optic disc and cup segmentation is difficult due to low contrast and vessel interference.
• Accurate segmentation is crucial for reliable glaucoma diagnosis and management.

Purpose of the Study:

• To present a novel automated glaucoma detection framework that utilizes the optic rim structure as a primary biomarker.
• To overcome the limitations of conventional methods by excluding reliance on CDR and RDR.

Main Methods:

• A Teacher-Student Generative Adversarial Network (TS-GAN) was developed for precise optic cup and disc segmentation.
• The Teacher model employed an attention-based CNN encoder-decoder, while the Student model used Expectation Maximization for enhanced segmentation.
• A rim generator created the optic rim from segmented structures, feeding into a SqueezeNet for glaucoma classification.

Main Results:

• The TS-GAN effectively addressed mode collapse, achieving superior segmentation accuracy compared to existing GANs.
• The complete framework demonstrated higher segmentation and glaucoma detection accuracy on diverse fundus image datasets, including private data.
• The proposed method proved effective for early glaucoma detection, offering enhanced reliability.

Conclusions:

• The novel framework provides a robust and accurate automated tool for glaucoma detection, focusing on the optic rim.
• This approach offers a reliable alternative to traditional methods, aiding ophthalmologists in efficient glaucoma management and vision loss prevention.