Evolutionary Compression of Deep Neural Networks for Biomedical Image Segmentation

Area of Science:

• Biomedical image analysis
• Deep learning
• Computational neuroscience

Background:

• Deep neural networks (DNNs) excel at biomedical image segmentation but are often too large for practical deployment.
• High parameter counts hinder the use of DNNs on real-time platforms and portable devices.

Purpose of the Study:

• To propose an evolutionary compression method (ECDNN) for automatically discovering efficient DNN architectures for biomedical image segmentation.
• To enable simultaneous optimization of network performance and parameter reduction.

Main Methods:

• Developed an evolutionary compression method (ECDNN) to automatically search for efficient DNN architectures.
• Implemented novel genetic operators for identifying and pruning less important filters, including those in feature map concatenation layers.
• Optimized network loss and parameter count concurrently, generating Pareto-optimal solutions.

Main Results:

• ECDNN successfully compressed DNNs for retinal vessel and neuronal membrane segmentation without retraining.
• The method discovered efficient architectures that maintained or improved performance.
• Experimental results demonstrated the superiority of ECDNN compared to state-of-the-art methods.

Conclusions:

• ECDNN provides an effective approach for creating lightweight and efficient DNNs for biomedical image segmentation.
• The method facilitates the deployment of advanced segmentation models on resource-constrained devices.
• ECDNN offers a flexible way to balance segmentation performance and model complexity.