Optimized-Unet: Novel Algorithm for Parapapillary Atrophy Segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Myopia prevalence is rising globally, particularly in China's youth.
• High myopia can lead to severe visual impairment and blindness.
• Parapapillary atrophy (PPA) is a key indicator of high myopia and requires precise segmentation for diagnosis.

Purpose of the Study:

• To develop an advanced deep learning model for accurate segmentation of parapapillary atrophy (PPA) in high myopia patients.
• To enhance the diagnostic capabilities for high myopia by improving PPA segmentation accuracy.

Main Methods:

• An optimized Unet (OT-Unet) architecture was proposed, incorporating pre-trained backbones (VGG, ResNet, Res2Net).
• Novel modules including edge attention, parallel partial decoder, and reverse attention were integrated to refine feature extraction and integration.
• An augmented loss function was introduced to improve the segmentation of complex lesion areas.

Main Results:

• The OT-Unet model achieved a high Area Under Curve (AUC) of 0.9235 on a dataset of 360 images.
• This represents a significant improvement compared to the original Unet, which achieved an AUC of 0.7917.
• The integrated modules effectively extracted contour information and combined multi-scale and hierarchical features.

Conclusions:

• The proposed OT-Unet demonstrates superior performance in PPA segmentation compared to the standard Unet.
• Accurate PPA segmentation using OT-Unet is crucial for the diagnosis and management of high myopia.
• This AI-driven approach holds promise for improving ophthalmological diagnostics and patient outcomes.