Synergistically segmenting choroidal layer and vessel using deep learning for choroid structure analysis.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computational Biology

Background:

• The choroid's vascular structure is vital for retinal physiology and ocular disease.
• Existing methods struggle to differentiate inner choroidal structures like vessels and stroma.

Purpose of the Study:

• To develop a synergistic segmentation pipeline for choroid layer and choroid vessel analysis.
• To introduce the choroidal U-shape network (CUNet) for improved choroidal segmentation.

Main Methods:

• Proposed a multi-task learning strategy using the CUNet architecture.
• Developed an adaptive multi-task segmentation loss to balance dual-task performance.
• Implemented pixel-wise classification for choroid layer and vessel segmentation.

Main Results:

• The CUNet pipeline demonstrated high performance with a 4% higher dice score.
• Achieved reduced computational complexity with an 18.85 M lower model size.
• The strategy showed strong generalization capabilities for both choroid layer and vessel segmentation.

Conclusions:

• The proposed CUNet pipeline offers a significant advancement in choroidal segmentation.
• The method's high performance and efficiency suggest considerable clinical potential for diagnosing and managing eye diseases.