A weak edge estimation based multi-task neural network for OCT segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical Coherence Tomography (OCT) provides high-resolution fundus images crucial for retinal health analysis, diagnosis, and treatment.
• Deep learning methods are increasingly used for fundus OCT image segmentation but struggle with weak edge sensitivity and data scarcity leading to overfitting.

Purpose of the Study:

• To address the limitations of current deep learning methods in fundus OCT image segmentation.
• To enhance the preservation of weak edge details and mitigate overfitting caused by limited annotated medical data.

Main Methods:

• Introduced the Multi-Task Attention Mechanism Network with Pruning (MTAMNP), featuring segmentation and boundary regression branches.
• Employed an adaptive weighted loss function based on Truncated Signed Distance Function (TSDF) in the boundary regression branch.
• Utilized a Spatial Attention Based Dual-Branch Information Fusion Block and a structured pruning method based on channel attention.

Main Results:

• The MTAMNP method demonstrated superior performance compared to state-of-the-art segmentation networks.
• Achieved Dice scores of 84.09% on the HCMS dataset and 93.84% on the Duke dataset.
• The structured pruning method effectively reduced parameter count and prevented overfitting while maintaining segmentation accuracy.

Conclusions:

• The proposed MTAMNP effectively tackles challenges in fundus OCT image segmentation, particularly weak edge preservation and overfitting.
• The dual-branch architecture and pruning strategy offer a robust solution for accurate and efficient medical image analysis.
• This advancement holds promise for improved automated diagnosis and treatment planning in ophthalmology.