Vison transformer adapter-based hyperbolic embeddings for multi-lesion segmentation in diabetic retinopathy

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) is a leading cause of global blindness.
• Early detection and treatment are vital for preventing vision loss.
• Automated diagnosis using deep learning shows promise for DR segmentation.

Purpose of the Study:

• To propose a novel Transformer-based model for enhanced Diabetic Retinopathy segmentation.
• To integrate hyperbolic embeddings and a spatial prior module into a Vision Transformer for improved DR diagnosis.

Main Methods:

• Developed a Vision Transformer encoder enhanced with a spatial prior module for image convolution and feature continuity.
• Utilized hyperbolic embeddings for pixel-level classification of feature matrices.
• Incorporated spatial feature injector and extractor for feature interaction processing.

Main Results:

• The proposed model demonstrated superior performance compared to existing DR segmentation models on public datasets.
• Hyperbolic embeddings effectively captured geometric structures in feature matrices, improving segmentation accuracy.
• The spatial prior module enhanced feature continuity, aiding lesion differentiation.

Conclusions:

• The novel Transformer-based model with hyperbolic embeddings and a spatial prior module significantly improves DR segmentation accuracy.
• This approach offers potential for accurate and rapid automated DR diagnosis in clinical settings.
• Future work may extend this model to other medical imaging tasks and clinical validation.