Genetic U-Net

Genetic U-Net applies an improved genetic algorithm to automatically optimize U-Net–style convolutional neural network architectures for accurate and parameter-efficient retinal vessel segmentation from fundus images.

Key Features:

• Automated Architecture Design: Uses an improved genetic algorithm to explore a condensed, flexible search space based on a U-shaped encoder-decoder (U-Net) structure to generate CNN architectures.
• Parameter Efficiency: Produces architectures with substantially fewer parameters, using less than 1% of the parameter count of the original U-Net and fewer than other state-of-the-art models.
• Enhanced Performance: Identifies effective operations and network patterns via exhaustive search within the defined architecture space to improve retinal vessel segmentation accuracy.
• Reduced Overfitting Risk: Lowers model complexity and parameter count to mitigate overfitting and improve robustness across diverse datasets.

Scientific Applications:

• Retinal vessel segmentation: Generates accurate vessel segmentations from fundus images for quantitative vascular analysis.
• Ophthalmic disease diagnosis and monitoring: Supports diagnosis and monitoring of ocular diseases such as diabetic retinopathy and glaucoma through vessel-based image biomarkers.
• Clinical research and diagnostics: Enables parameter-efficient models suitable for medical imaging research and diagnostic workflows.

Methodology:

Search space definition based on a U-shaped encoder-decoder framework; improved genetic algorithm optimization to iteratively explore and select candidate architectures; performance evaluation of resulting architectures on retinal vessel segmentation tasks with focus on accuracy and parameter efficiency.