JOINEDTrans: Prior guided multi-task transformer for joint optic disc/cup segmentation and fovea detection

Area of Science:

• Ophthalmology
• Medical Image Analysis
• Computer Vision

Background:

• Deep learning models have advanced retinal landmark analysis (optic disc/cup, fovea).
• Ophthalmic lesions and poor image quality challenge automatic segmentation and detection.
• Existing methods often address landmarks individually without leveraging prior information.

Purpose of the Study:

• To introduce JOINEDTrans, a prior-guided multi-task transformer framework for joint optic disc/cup segmentation and fovea detection.
• To improve the robustness of retinal landmark analysis against lesions and image quality issues.
• To incorporate spatial priors, specifically vessel information, into the multi-task learning framework.

Main Methods:

• A multi-task transformer framework (JOINEDTrans) with segmentation and detection branches.
• Incorporation of spatial priors through an encoder pre-trained on vessel segmentation.
• A two-stage approach: coarse segmentation/localization followed by fine refinement using region cropping.

Main Results:

• JOINEDTrans effectively combines spatial features to mitigate distortions from lesions and imaging artifacts.
• The framework demonstrated superior performance compared to state-of-the-art methods.
• Experiments were conducted on the GAMMA, REFUGE, and PALM fundus image datasets.

Conclusions:

• JOINEDTrans offers a robust and accurate solution for joint optic disc/cup segmentation and fovea detection.
• The incorporation of spatial priors significantly enhances performance in challenging conditions.
• The proposed method advances automated analysis of retinal landmarks in ophthalmology.