Evaluating UNet Performance and Showcasing the Computational Benefits of SwinUNet for Liver Tumor Segmentation

Liver cancer is a significant global health challenge, causing over 830,000 deaths each year. This research evaluates Unet’s performance in liver tumor segmentation, noting its effectiveness in detecting larger lesions but also its limitations with complex structures. Unet’s design, relying on convolutional layers, makes it computationally intensive for high-resolution images and weak in capturing long-range dependencies, leading to less accurate segmentation in some cases. To address these issues, this study introduces SwinUNet, which integrates Swin Transformers with Unet. SwinUNet efficiently captures multiscale features and reduces computational complexity through feature pyramids and a shifted-window self-attention mechanism. This approach enhances segmentation accuracy in challenging cases and improves processing efficiency, making it suitable for a variety of clinical applications. By extending the functionality of Unet, SwinUNet shows promise in advancing liver tumor diagnosis and treatment strategies.