Abstract

Accurate boundary segmentation of volumetric images is a critical task for image-guided diagnosis and computer-assisted intervention. It is challenging to address the boundary confusion with explicit constraints. Existing methods of refining boundaries overemphasize the slender structure while overlooking the dynamic interactions between boundaries and neighboring regions. In this paper, we reconceptualize the mechanism of boundary generation via introducing Pushing and Pulling interactions, then propose a unified network termed PP-Net to model shape characteristics of the confused boundary region. Specifically, we first propose the semantic difference module (SDM) from the pushing branch to drive the boundary towards the ground truth under diffusion guidance. Additionally, the class clustering module (CCM) from the pulling branch is introduced to stretch the intersected boundary along the opposite direction. Thus, pushing and pulling branches will furnish two adversarial forces to enhance representation capabilities for the faint boundary. Experiments are conducted on four public datasets and one in-house dataset plagued by boundary confusion. The results demonstrate the superiority of PP-Net over other segmentation networks, especially on the evaluation metrics of Hausdorff Distance and Average Symmetric Surface Distance. Besides, SDM and CCM can serve as plug-and-play modules to enhance classic U-shape baseline models, including recent SAM-based foundation models. Source codes are available at https://github.com/EndoluminalSurgicalVision-IMR/PnPNet.

Fig. 1. Some typical cases plagued by boundary confusion. (a) Faint boundaries between the pulmonary lobes. (b) Boundary delineation between non-enhancing tumor core and peritumoral edema. (c) Multi-class vertebrae segmentation. (d) The boundary segmentation between the left atrium and left atrial appendage. (a), (b), (d): From top to bottom: image/ prediction by 3D UNet/ground truth. (c): From left to right: image/ prediction by 3D UNet/ground truth.

Fig. 2. Two adversarial dynamics acting on the inter-class boundary. For simplicity, we illustrate the relation description between two regions. The boundary interface is stretched opposite to the boundary GT by the pulling force, which arises from clustered class centers. In the meanwhile, the boundary is driven towards the GT position via the diffusion-driven pushing force. During the interactive process, two adversarial forces will diminish gradually, then achieve an equilibrium for a precise boundary delineation.

https://doi.org/10.1016/j.media.2026.103961