ViT ‐Swin‐ UNet : A Transformer‐Based Framework for Liver Tumor Analysis in CT Images

Precise classification and segmentation are essential for detecting life‐threatening conditions such as liver cancer. Early detection and treatment planning will improve the survival rates of the patient. 3D structures of liver cancer can be obtained through Computed Tomography (CT). But analyzing these data needs manual analysis, which remains laborious and skill oriented. From these problem statements, this experiment proposes a hybrid deep‐learning framework for slice‐wise liver tumor classification, which combines Vision Transformer (ViT) and Swin‐UNet and is known as ViT‐Swin‐UNet. The pipeline transforms 3D scans into 2D slices with annotated segmentation masks (tumor present/absent). The ViT is trained to achieve 95.4% accuracy, 0.975 AUC, and 95.38% F1‐score for liver tumor identification. The Swin‐UNet model achieved a test Dice score of 0.8957 in tumor segmentation, successfully locating tumors with different sizes and texture. This dual capability of detecting whether a slice contains a tumor and precisely localizing its position addresses a critical gap in automated diagnostics, offering radiologists enhanced efficiency along with improved interpretability. The system automates tumor detection and prioritizes high‐risk slices, reducing the radiologists' manual review burden, speeding up and improving diagnosis without compromising precision, bridging AI efficiency with clinical practicality in resource‐constrained oncology workflows.