C2FTFNet: Coarse-to-fine transformer network for joint optic disc and cup segmentation

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a primary cause of global blindness, necessitating early detection through methods like Cup-to-Disk Ratio (CDR) evaluation.
• Deep learning excels in optic disk (OD) and optic cup (OC) segmentation for computer-aided diagnosis (CAD) systems.
• Clinical data complexity can limit current deep learning techniques for glaucoma screening.

Purpose of the Study:

• To introduce an novel Coarse-to-Fine Transformer Network (C2FTFNet) for precise joint segmentation of OD and OC.
• To enhance glaucoma screening by improving the accuracy of OD and OC segmentation.
• To address limitations in existing deep learning models when dealing with complex clinical data.

Main Methods:

• A two-stage C2FTFNet: coarse stage uses U-Net and Circular Hough Transform (CHT) for Region of Interest (ROI) segmentation, and fine stage employs TransUnet3+ for accurate OC and OD extraction.
• Incorporation of a Transformer module to capture long-range dependencies and global information, overcoming convolutional limitations.
• A Multi-Scale Dense Skip Connection (MSDC) module to effectively fuse multi-level features and reduce semantic gaps.

Main Results:

• The C2FTFNet demonstrated superior performance in OD and OC segmentation compared to state-of-the-art methods.
• Experiments on DRIONS-DB, Drishti-GS, and REFUGE datasets validated the model's effectiveness.
• The proposed architecture successfully segmented critical ocular structures for glaucoma assessment.

Conclusions:

• C2FTFNet offers a robust and effective approach for glaucoma screening via accurate OD and OC segmentation.
• The model's ability to handle complex clinical data signifies a significant advancement in automated ophthalmic diagnostics.
• This deep learning framework holds promise for improving early detection and prevention of vision loss due to glaucoma.