Structural Priors Guided Network for the Corneal Endothelial Cell Segmentation.

Area of Science:

• Medical Imaging Analysis
• Deep Learning for Computer Vision
• Ophthalmology

Background:

• Segmenting blurred cell boundaries in corneal endothelium images is difficult.
• Current deep learning methods neglect cell structure, leading to discontinuous segmentation.
• This impacts the accuracy of clinical parameter estimation.

Purpose of the Study:

• To develop a novel network, SPG-Net, for accurate corneal endothelium cell segmentation.
• To address the limitations of existing methods by incorporating structural prior knowledge.
• To improve the continuity and accuracy of cell boundary segmentation.

Main Methods:

• Utilized a hybrid transformer-convolution backbone for enhanced global context capture.
• Implemented Feature Enhancement (FE) and Local Affinity-based Feature Fusion (LAFF) modules for feature representation and structural information propagation.
• Introduced a joint loss function combining cross-entropy and Structure Similarity Index Measure (SSIM) for pixel and structure level supervision.

Main Results:

• SPG-Net demonstrated superior performance compared to state-of-the-art methods on four corneal endothelial datasets.
• The proposed method effectively reduced discontinuous cell boundary segmentation.
• SPG-Net achieved a good balance between pixel-wise accuracy and structure preservation.

Conclusions:

• SPG-Net successfully overcomes the challenge of segmenting blurred cell boundaries in corneal endothelium images.
• The method shows good agreement and correlation in clinical parameter estimation compared to ground truth.
• SPG-Net offers a promising solution for improving diagnostic accuracy in ophthalmology.