Medical Image Segmentation Using Automatic Optimized U-Net Architecture Based on Genetic Algorithm.

Area of Science:

• Computer Vision
• Medical Imaging
• Artificial Intelligence

Background:

• Biomedical image segmentation is vital for clinical decisions, driving research in computer vision.
• Deep learning methods achieve high performance but demand significant expert knowledge and computational resources.
• Existing models often have complex architectures with numerous parameters.

Purpose of the Study:

• To propose GA-UNet, a novel approach using genetic algorithms to automatically design efficient U-Net architectures.
• To minimize architectural complexity and parameter count in deep learning models for biomedical image segmentation.
• To address the computational demands and expert knowledge requirements of current segmentation methods.

Main Methods:

• Development of GA-UNet, employing genetic algorithms for automated neural network architecture design.
• Evaluation of GA-UNet on lung, cell nuclei (DSB 2018), and liver image segmentation datasets.
• Comparison of GA-UNet's performance and parameter count against the original U-Net model.

Main Results:

• GA-UNet achieved competitive accuracy: 98.78% (lung), 95.96% (cell nuclei), 98.58% (liver).
• The model utilizes significantly fewer parameters: 0.24% (lung), 0.48% (cell nuclei), 0.67% (liver) of original U-Net.
• Demonstrated superior architectural efficiency compared to the standard U-Net.

Conclusions:

• GA-UNet offers a viable solution for efficient biomedical image segmentation.
• The approach is suitable for resource-limited environments and real-world applications requiring fast inference.
• Automated architecture design via genetic algorithms can yield high-performing, less complex deep learning models.