Joint disc and cup segmentation based on recurrent fully convolutional network

Area of Science:

• Ophthalmology
• Medical Image Analysis
• Computer Vision

Background:

• Accurate segmentation of the optic disc (OD) and optic cup (OC) is crucial for diagnosing various eye conditions from fundus images.
• Existing Fully Convolutional Network (FCN) methods struggle with dense prediction tasks due to information loss from consecutive convolution and pooling operations.

Purpose of the Study:

• To develop an automated method for joint segmentation of the optic disc and optic cup in fundus images.
• To improve the retention of detailed spatial and subtle edge information lost in traditional FCNs.

Main Methods:

• Proposed Recurrent Fully Convolution Network (RFC-Net) incorporating a multi-scale input layer, recurrent convolutional layers, multiple output layers, and polar transformation.
• Utilized four novel recurrent units for feature accumulation and enhanced feature representation.
• Employed a multiple output cross-entropy loss function and polar transformation to balance cup ratio and improve segmentation accuracy.

Main Results:

• RFC-Net demonstrated superior segmentation performance compared to original FCN and other state-of-the-art methods on the DRISHTI-GS1 dataset.
• The proposed network effectively minimizes spatial information loss, capturing high-level and subtle edge details.
• Achieved better segmentation accuracy and generalization capabilities in fundus image analysis.

Conclusions:

• RFC-Net offers a significant advancement in automated optic disc and optic cup segmentation.
• The method's ability to preserve spatial details and handle complex features makes it highly effective for medical image analysis.
• This approach holds promise for improved diagnostic tools in ophthalmology.