Retinal Vessel Segmentation using Round-wise Features Aggregation on Bracket-shaped Convolutional Neural Networks.

Area of Science:

• Medical Image Processing
• Deep Learning
• Computer Vision

Background:

• Retinal blood vessel segmentation is crucial for diagnosing eye diseases.
• Deep learning methods often require extensive data augmentation (patches) and computational resources due to imbalanced pixel data.
• Existing approaches struggle with the diverse and irregular nature of retinal vasculature.

Purpose of the Study:

• To propose a novel deep learning architecture, Round-wise Features Aggregation on Bracket-shaped convolutional neural networks (RFA-BNet), for retinal blood vessel segmentation.
• To eliminate the need for image patch augmentation, thereby reducing computational costs.
• To effectively handle the complex and varied representations of retinal blood vessels.

Main Methods:

• Utilized a bracket-shaped convolutional neural network (BNet) architecture.
• Employed round-wise feature aggregation, extracting and combining features from different stages of a pretrained backbone network.
• Focused on exploiting middle-scale features iteratively within the decoder.
• Aggregated highest-resolution decoded maps from each round to capture fine details.

Main Results:

• Achieved high performance on the DRIVE dataset.
• Demonstrated effectiveness in segmenting thin and small vessels.
• Reported a sensitivity of 0.7932, specificity of 0.9741, accuracy of 0.9511, and AUROC of 0.9732.

Conclusions:

• RFA-BNet offers an efficient and effective solution for retinal blood vessel segmentation.
• The proposed method successfully addresses the challenges of imbalanced data and complex vessel structures without patch augmentation.
• This approach shows significant potential for improving automated retinal image analysis.