Liver cancer poses a significant global health challenge, necessitating precise tumor segmentation in CT scans for diagnosis and treatment. While deep learning models like U-Net and Vision Transformers show promise, their computational demands hinder edge deployment. To address this gap, we propose an optimized Swin-UNet framework enhanced by the Search and Rescue (SAR) algorithm, enabling real-time edge computing without compromising the model’s performance. This work proposes a hybrid objective function with quadratic penalties for model compression and area under the curve(AUC). The models are trained using focal AUC loss to mitigate class imbalance. Evaluations on 3DIRCADb, LiTS, and MSD datasets show state-of-the-art performance, with Dice scores of 94.78%, 89.06%, and 88.95%, respectively, and an 80.3% parameter reduction versus baselines. The solution achieves efficient segmentation on edge devices (e.g., Jetson Nano), with a Volume Overlap Error of 1.73% (MSD) and Relative Volume Difference of 0.23% (3DIRCADb), outperforming existing methods. This work advances memory-efficient deep learning for clinical deployment, enabling AI-driven diagnostics in low-resource settings.
Liver cancer remains a critical global health challenge, with accurate segmentation of tumors in CT scans being vital for diagnosis. While deep learning models like U-Net and Vision Transformers offer promising results, their high computational demands hinder deployment on resource-constrained edge devices. This study introduces an optimized Swin-UNet model for efficient liver tumor segmentation, leveraging a Search and Rescue (SAR) algorithm to balance model size and accuracy. Key innovations include a quadratic penalty-based objective function for joint optimization of AUC and model compactness and a focal AUC loss to address class imbalance. Evaluated across three datasets (3DIRCADb, LiTS, MSD), the proposed SAR-Swin-UNet achieves superior performance, with Dice scores of 94.78%, 89.06% and 88.95%, respectively, while reducing the model size by up to 80.3% compared to unoptimized counterpart. The approach enables real-time, energy-efficient segmentation on edge devices like Jetson Nano, addressing challenges of data security and computational costs. Results demonstrate significant improvements over state-of-the-art methods with a minimum Volume Overlap Error of (1.73%) in MSD dataset and Relative Volume Difference of (0.23%) in 3DIRCADb dataset, highlighting its clinical applicability for precise, low-resource settings. This work bridges the gap between high-accuracy segmentation and practical deployment in under-resourced healthcare environments.