RTNet: Relation Transformer Network for Diabetic Retinopathy Multi-Lesion Segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Diabetic retinopathy (DR) lesion segmentation aids ophthalmologists in diagnosis.
• Prior research often overlooked pathological associations between DR lesions and vascular structures.
• Understanding these relationships is crucial for accurate lesion identification.

Purpose of the Study:

• To develop an automated system for segmenting diabetic retinopathy lesions.
• To incorporate pathological associations and vascular information into the segmentation process.
• To improve the accuracy and robustness of DR lesion detection.

Main Methods:

• Proposed a relation transformer block (RTB) with self-attention and cross-attention mechanisms.
• Integrated vascular information to resolve ambiguity in lesion detection.
• Introduced a global transformer block (GTB) to preserve fine details of small lesions.
• Developed a dual-branch network for simultaneous segmentation of four lesion types.

Main Results:

• The proposed network achieved competitive performance on IDRiD and DDR datasets.
• Demonstrated the superiority of incorporating pathological associations and vascular information.
• Successfully segmented multiple DR lesion types simultaneously.

Conclusions:

• The novel approach effectively utilizes pathological associations for DR lesion segmentation.
• The method shows significant potential in assisting ophthalmologists with DR diagnosis.
• Future work can explore further integration of clinical data for enhanced diagnostic tools.