.png){kind=icon}
.png){kind=icon}
Return
Adversarial-consistency enhanced implicit segmentation field for weakly supervised 3D cardiac image segmentation
.png){kind=icon}
PRE
AI
W
J
高
J
J
伍
C
C
Z
刘
DOI：10.1016/j.media.2026.104094.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• We extend the inference path from the pixel space to the coordinate space. In coordinate space, we construct an implicit function to integrate the relative topological relationships, thus enhancing the discriminability between cardiac chambers. • We propose an adaptive adversarial consistency to tackle label sparsity. It leverages adversarial learning to estimate the underlying distribution, thereby providing an adaptive sampling range for consistency regularization. • We design a label augmentation algorithm to prevent mode collapse in adversarial learning. It distills initial labels from a foundation model and filters noise through collaborative optimization between the superpixels and the implicit function. • Extensive experiments on four types of scribble-level annotations show that ACISF achieves outstanding performance over ten state-of-the-art methods.
Keywords:
Implicit function
Adversarial consistency
Weakly supervised segmentation
Cardiac image segmentation
Coordinate space
Journal
IF:
11.8
Papers:
3.7K
Citations:
2.4W
Organization
Cited Papers