Lightweight Evolving U-Net for Next-Generation Biomedical Imaging

Background/Objectives: Accurate and efficient segmentation of cell nuclei in biomedical images is critical for a wide range of clinical and research applications, including cancer diagnostics, histopathological analysis, and therapeutic monitoring. Although U-Net and its variants have achieved notable success in medical image segmentation, challenges persist in balancing segmentation accuracy with computational efficiency, especially when dealing with large-scale datasets and resource-limited clinical settings. This study aims to develop a lightweight and scalable U-Net-based architecture that enhances segmentation performance while substantially reducing computational overhead. Methods: We propose a novel evolving U-Net architecture that integrates multi-scale feature extraction, depthwise separable convolutions, residual connections, and attention mechanisms to improve segmentation robustness across diverse imaging conditions. Additionally, we incorporate channel reduction and expansion strategies inspired by ShuffleNet to minimize model parameters without sacrificing precision. The model performance was extensively validated using the 2018 Data Science Bowl dataset. Results: Experimental evaluation demonstrates that the proposed model achieves a Dice Similarity Coefficient (DSC) of 0.95 and an accuracy of 0.94, surpassing state-of-the-art benchmarks. The model effectively delineates complex and overlapping nuclei structures with high fidelity, while maintaining computational efficiency suitable for real-time applications. Conclusions: The proposed lightweight U-Net variant offers a scalable and adaptable solution for biomedical image segmentation tasks. Its strong performance in both accuracy and efficiency highlights its potential for deployment in clinical diagnostics and large-scale biological research, paving the way for real-time and resource-conscious imaging solutions.