Lightweight Evolving U-Net for Next-Generation Biomedical Imaging.

Area of Science:

• Biomedical image analysis
• Computational pathology
• Deep learning for medical imaging

Background:

• Accurate cell nuclei segmentation is vital for cancer diagnostics and research.
• Existing U-Net models face challenges in balancing accuracy and computational efficiency.
• Large datasets and limited clinical resources necessitate efficient segmentation solutions.

Purpose of the Study:

• To develop a lightweight and scalable U-Net architecture for enhanced biomedical image segmentation.
• To improve segmentation performance while reducing computational overhead.
• To address the need for efficient solutions in resource-limited clinical settings.

Main Methods:

• Proposed a novel evolving U-Net architecture integrating multi-scale feature extraction, depthwise separable convolutions, residual connections, and attention mechanisms.
• Incorporated channel reduction and expansion strategies (inspired by ShuffleNet) to minimize parameters.
• Validated model performance using the 2018 Data Science Bowl dataset.

Main Results:

• Achieved a Dice Similarity Coefficient (DSC) of 0.95 and an accuracy of 0.94, outperforming state-of-the-art benchmarks.
• Demonstrated high-fidelity delineation of complex and overlapping nuclei.
• Maintained computational efficiency suitable for real-time applications.

Conclusions:

• The lightweight U-Net variant provides a scalable and adaptable solution for biomedical image segmentation.
• Strong performance in accuracy and efficiency supports potential clinical and research deployment.
• Paves the way for real-time, resource-conscious imaging solutions in diagnostics and biological research.