Exploiting multi-granularity visual features for retinal layer segmentation in human eyes.

Area of Science:

• Medical imaging
• Computer vision
• Ophthalmology

Background:

• Accurate retinal layer segmentation is crucial for early ophthalmic disease detection.
• Existing segmentation algorithms often lack multi-granularity feature exploitation and dataset availability.
• Deep learning solutions require accessible, high-quality datasets for robust model training.

Purpose of the Study:

• To develop an advanced end-to-end deep learning network for precise retinal layer segmentation.
• To introduce a novel dataset (NR206) for retinal image segmentation research.
• To enhance feature retention and multi-scale analysis in segmentation models.

Main Methods:

• Proposed a novel end-to-end retinal layer segmentation network utilizing the ConvNeXt architecture.
• Incorporated a depth-efficient attention module and multi-scale structures to preserve feature map details.
• Developed and released the NR206 dataset comprising 206 healthy human retinal images for semantic segmentation.

Main Results:

• The proposed network achieved superior performance on the NR206 dataset, with an average Dice score of 91.3% and mIoU of 84.4%.
• The model demonstrated state-of-the-art performance on glaucoma and diabetic macular edema (DME) datasets.
• The NR206 dataset is user-friendly, requiring no additional transcoding.

Conclusions:

• The novel ConvNeXt-based network offers a significant advancement in retinal layer segmentation accuracy.
• The publicly available NR206 dataset and source code will facilitate further research in deep learning for ophthalmology.
• The model's effectiveness across multiple ophthalmic datasets highlights its versatility and potential clinical applicability.