Adversarial-consistency enhanced implicit segmentation field for weakly supervised 3D cardiac image segmentation.

Automatic acquisition of 3D cardiac anatomical structures plays a crucial role in the diagnosis of heart diseases. However, most existing automatic methods rely on costly pixel-level dense annotations. To reduce labeling costs, scribble-level supervised learning provides a promising solution. However, accurate cardiac segmentation with scribble-level supervision remains challenging due to intrinsic similarity and label sparsity. To address these issues, we propose the adversarial-consistency enhanced implicit segmentation field (ACISF). The ACISF extends the inference path from pixel space to coordinate space and construct an implicit function to directly model semantic information within coordinate space. In this way, the implicit function integrates discriminative cues provided by the relative topological relationships between cardiac chambers in the coordinate space, mitigating the challenge posed by intrinsic similarity. Building on this, we further propose adaptive adversarial consistency (AAC) to tackle label sparsity. AAC leverages adversarial learning to approximate the underlying distribution within the neighborhood of sampled coordinates. This distribution provides an adaptive sampling range for consistency regularization, avoiding the unreliability of manually defined sampling ranges. Finally, we conduct extensive experiments on four types of scribble-level annotations. Results on 683 patients demonstrate that our method outperforms ten state-of-the-art methods.