Cup and Disc Segmentation in Smartphone Handheld Ophthalmoscope Images with a Composite Backbone and Double Decoder...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma diagnosis relies on optic disc and cup examination in fundus images.
• Current screening methods are costly and impractical for primary care.
• Deep neural networks can assist in segmenting optic nerve structures.

Purpose of the Study:

• To enhance morphological biomarkers of the optic disc and cup using smartphone-acquired fundus images.
• To develop a deep neural network for improved segmentation accuracy.
• To introduce a novel approach for combining loss weights during training.

Main Methods:

• A deep neural network architecture combining two backbones and a dual decoder was employed.
• The methodology focused on segmenting optic disc and cup structures.
• Loss weights were combined in a new way during the training process.

Main Results:

• The developed models achieved high segmentation performance.
• Dice scores reached 95.92% for the optic disc and 85.30% for the optic cup.
• IoU scores reached 92.22% for the optic disc and 75.68% for the optic cup on the BrG dataset.

Conclusions:

• The proposed deep neural network architecture shows promise for fundus image segmentation.
• This approach can aid in the early diagnosis of glaucoma.
• Smartphone-based imaging combined with AI offers a feasible screening solution.