Sli2Vol+: Segmenting 3D Medical Images Based on an Object Estimation Guided Correspondence Flow Network

Area of Science:

• Medical imaging
• Artificial intelligence
• Computer vision

Background:

• Deep learning (DL) excels in medical image segmentation but requires extensive annotated data, which is costly and difficult to obtain for 3D volumes.
• Existing mask propagation DL methods reduce annotation burden but suffer from error accumulation and struggle with discontinuities between slices.

Purpose of the Study:

• To introduce Sli2Vol+, a novel self-supervised framework (SSF) for 3D medical image segmentation using only a single annotated slice per volume.
• To address limitations of previous methods, specifically error accumulation and handling of discontinuities.

Main Methods:

• Sli2Vol+ generates pseudo-labels (PLs) by propagating an annotated 2D slice within a training volume.
• A novel Object Estimation Guided Correspondence Flow Network is developed for self-supervised learning of correspondences between slices and PLs.
• These learned correspondences are used in the test stage to propagate a single annotated slice for segmentation.

Main Results:

• The method demonstrates effectiveness across various medical image segmentation tasks and datasets.
• Sli2Vol+ shows improved generalizability across different organs, modalities, and imaging modes.
• The approach successfully segments anatomical structures with significantly reduced annotation effort.

Conclusions:

• Sli2Vol+ offers an effective solution for 3D medical image segmentation with minimal annotation requirements.
• The proposed SSF overcomes limitations of prior mask propagation techniques, enhancing reliability and accuracy.
• This method holds potential for broader application in medical image analysis, facilitating efficient segmentation workflows.