U-NTCA: nnUNet and nested transformer with channel attention for corneal cell segmentation

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Automatic segmentation of corneal stromal cells is crucial for diagnosing eye conditions like viral infections and keratoconus from confocal microscopy images.
• Challenges in segmentation include uneven illumination and vascular occlusion, leading to inaccurate results.

Purpose of the Study:

• To develop a novel deep learning architecture for accurate corneal stromal cell segmentation, addressing limitations of existing methods.
• To improve the detection of abnormal cell morphology in challenging microscopy images.

Main Methods:

• Proposed a novel nnUNet and nested Transformer-based network (U-NTCA) with dual high-order channel attention.
• Incorporated recursive feature transmission and cross-layer interaction for enhanced contextual understanding.
• Utilized a novel attention channel (gConv) for higher-order local context interaction and multi-scale feature integration.

Main Results:

• The U-NTCA model achieved a Dice score of 82.72% and an Area Under Curve (AUC) of 90.92% on a clinical dataset of 136 images.
• Demonstrated superior performance compared to the standard nnUNet architecture.

Conclusions:

• The proposed U-NTCA model offers a cost-effective and high-precision solution for corneal stromal cell segmentation.
• The method is particularly effective in challenging image scenarios, improving diagnostic accuracy for ophthalmologists.