[Segmentation of retinal image vessels based on fully convolutional network with depthwise separable convolution and...

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Context:

• Diabetic retinopathy and hypertension alter retinal blood vessel morphology, necessitating accurate segmentation for diagnosis.
• Quantitative analysis of retinal vasculature is crucial for early disease detection and monitoring.
• Existing segmentation methods may lack the precision required for subtle vascular changes.

Purpose:

• To develop an advanced Fully Convolutional Network (FCN) for precise retinal blood vessel segmentation.
• To enhance image contrast and expand training data using CLAHE, Gamma correction, and image patching.
• To improve feature representation and network performance through depthwise separable convolutions and channel weighting.

Summary:

• A novel FCN architecture incorporating depthwise separable convolutions and channel weighting is proposed for retinal vessel segmentation.
• Preprocessing steps include CLAHE and Gamma correction on the green channel, followed by image patching for data augmentation.
• The model was trained and validated on the DRIVE and STARE datasets, achieving high segmentation accuracy and AUC.

Impact:

• The proposed method demonstrates superior feature resolution and segmentation performance compared to existing techniques.
• Achieved high accuracy (0.963) and AUC (0.983) on the DRIVE dataset, and (0.962, 0.983) on the STARE dataset.
• This advancement holds potential for more accurate and efficient clinical diagnosis of retinal vascular diseases.