ABSTRACT Liver tumor segmentation in computed tomography (CT) images is challenging due to diverse tumor morphologies, indistinct boundaries, and background noise, which hinder contextual understanding and boundary delineation. To address these issues, we propose DGT‐MSB‐Net, integrating a dynamically guided Transformer (DGT) with multiscale boundary supervision. The DGT module employs offset‐guided deformable sampling and efficient channel modeling to capture complex and irregular tumor structures, while a multiscale spatial‐channel attention (MSCA) mechanism leverages skip connections to fuse shallow boundary cues with deep semantic features. In addition, a boundary‐aware (BA) module introduces explicit boundary supervision to refine edge prediction and enhance segmentation accuracy. Experiments on LiTS2017 and 3DIRCADb datasets show that DGT‐MSB‐Net outperforms existing methods, achieving consistent improvements in Dice, Precision, and Recall, validating its effectiveness and robustness. By combining dynamic feature modeling with BA strategies, DGT‐MSB‐Net effectively addresses the challenges of liver tumor segmentation and demonstrates strong potential for clinical decision support and automated analysis.
Yang et al. (Fri,) studied this question.