FR-UNet

FR-UNet segments retinal vessels and coronary angiographs using a full-resolution convolutional network combined with a dual-threshold iterative algorithm to detect thin, low-contrast vascular structures.

Key Features:

• Full-Resolution Network (FR-UNet): Maintains full image resolution throughout the segmentation process via a full-resolution architecture.
• Multiresolution Convolution Interactive Mechanism: Expands horizontally and vertically to preserve multiresolution spatial information across stages.
• Feature Aggregation Module: Integrates multiscale feature maps from adjacent stages to enrich high-level contextual information.
• Modified Residual Blocks: Continuously learn multiresolution representations to produce pixel-level accuracy prediction maps.
• Dual-Threshold Iterative Algorithm (DTI): Iteratively adjusts thresholds to refine segmentation and recover weak vessel signals.
• Weak Vessel Pixel Extraction: Identifies and extracts low-contrast vessel pixels to improve vessel connectivity.
• Sensitivity Enhancement: Increases detection sensitivity for thin and subtle vascular structures.

Scientific Applications:

• Retinal Vessel Datasets: Evaluated on DRIVE, CHASE_DB1, and STARE.
• Coronary Angiography Datasets: Evaluated on DCA1 and CHUAC.
• Benchmarking Metrics: Demonstrated superior sensitivity (Sen), area under the curve (AUC), F1 score, and intersection over union (IOU) while using fewer parameters than comparison models.

Methodology:

The method combines a full-resolution convolutional network with multiresolution convolution interactive mechanisms, feature aggregation modules and modified residual blocks to preserve spatial information and produce pixel-level prediction maps, together with a dual-threshold iterative algorithm that iteratively adjusts thresholds to extract weak vessel pixels and improve connectivity and sensitivity.