DynTransNet: Dynamic Transformer Network with multi-scale attention for liver cancer segmentation

Abstract

Introduction: Hepatocellular carcinoma (HCC), a predominant subtype of liver cancer, remains Q7 a major contributor to global cancer mortality. Accurate delineation of liver tumors in CT and MRI scans is critical for treatment planning and clinical decision-making. However, manual segmentation is time-consuming, errorprone, and inconsistent, necessitating reliable automated approaches. Methods: This study presents a novel U-shaped segmentation framework inspired by U-Net, designed to enhance accuracy and robustness. The encoder incorporates Dynamic Multi-Head Self-Attention (D-MSA) to capture both global and local spatial dependencies, while the decoder uses skip connections to preserve spatial detail. Additionally, a Feature Mix Module (FM-M) blends multiscale features, and a Residual Module (RM) refines feature representations and stabilizes training. The proposed framework addresses key challenges such as boundary precision, complex structural relationships, and dataset imbalance. Results: Experimental results demonstrate superior segmentation performance, achieving a mean Dice score of 86.12 on the ATLAS dataset and 93.12 on the LiTS dataset. Discussion: The proposed method offers a robust, efficient tool for liver tumor segmentation and holds strong potential to streamline diagnostic workflows and improve automated medical image analysis in clinical practice.