Sli2Vol: Annotate a 3D Volume from a Single Slice with Self-supervised Learning

Pak-Hei Yeung, Ana I. L. Namburete, Weidi Xie

Abstract

3D medical image segmentation is critical for diagnosis but often limited by the need for extensive labeled datasets and poor generalization across domains. We present Sli2Vol, a self-supervised framework that segments an entire 3D volume from a single user-annotated slice. By learning pixel-level correspondence between adjacent slices through an edge-based bottleneck, Sli2Vol achieves high accuracy and robust generalization across various organs and imaging modalities without requiring labeled training data.

Introduction

The "data bottleneck" in 3D medical imaging stems from the high cost of manual annotation. While fully automatic models like U-Nets are powerful, they require thousands of labeled volumes and often fail when applied to data from different scanners (domain shift). Semi-automatic tools offer a middle ground but frequently require tedious parameter tuning. Sli2Vol addresses these challenges by focusing on correspondence matching—learning how anatomical structures relate across slices rather than classifying specific organ types.

Methodology

The Sli2Vol framework consists of a self-supervised training phase and an inference pipeline designed for clinical efficiency.

Overview of the Sli2Vol framework, showing self-supervised training (top left), the Correspondence Flow Network (top right), and the inference process (bottom).

Self-Supervised Training

The model is trained on unlabeled 3D volumes. It learns to reconstruct a target slice $S_2$ by "weighting and copying" pixels from an adjacent slice $S_1$. This process generates an affinity matrix that captures spatial relationships. To prevent the network from relying on simple intensity matching, an Edge Profile Generator forces the model to focus on structural boundaries rather than raw pixel intensity.

Inference and Propagation

During inference, a user annotates a single slice. The learned affinity matrices propagate this mask through the volume. To combat "error drift," a Verification Module performs intensity-based checks, ensuring the propagated mask remains consistent with the original annotation's characteristics. The system is highly efficient, propagating masks at a speed of 2.27 slices per second.

Experiments and Results

Sli2Vol was evaluated across 8 public datasets covering 9 structures of interest (SOIs), including the liver, spleen, and heart, using both CT and MRI modalities.

Quantitative evaluation of Sli2Vol across multiple datasets and ablation studies.

Performance and Generalization

As shown in the results table, Sli2Vol achieves Dice scores exceeding 80 for most benchmarks. Notably, it demonstrates superior robustness in cross-domain scenarios. While fully supervised models often see performance degradation when tested on new scanners (dropping to scores like 56.0 in some cases), Sli2Vol's correspondence-based approach remains stable.

Ablation Studies

Ablation results highlight the importance of each component. The combination of the Correspondence Flow Network, Edge Profile Generator, and Verification Module (row 'j' in the table) yields the best mean results (78.2), significantly outperforming baseline optical flow (48.6) or standard supervised single-slice methods (60.4).

Conclusion

Sli2Vol provides a practical solution for 3D medical annotation, offering a balance between manual effort and automated precision. By prioritizing correspondence over classification, it achieves remarkable generalization across different organs and imaging sites. Future work will explore interactive refinements to handle highly discontinuous anatomical structures.

Summary updated

BibTeX
@inproceedings{yeung2021sli,
  author = {Pak-Hei Yeung and Ana I. L. Namburete and Weidi Xie},
  title = {{Sli2Vol: Annotate a 3D Volume from a Single Slice with Self-supervised Learning}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2021},
  url = {https://openreview.net/forum?id=YrPHEU54RS4u},
}
Figure