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

About this paper

Delin An, Pengfei Gu, Milan Sonka, Chaoli Wang, Danny Z. Chen

Sli2Vol+, a self-supervised mask propagation framework, enables 3D medical image segmentation using only a single 2D annotated slice per volume during training and testing. The method demonstrates improved segmentation accuracy, surpassing previous mask propagation methods by an average of 5.6 Dice score points on CT and MRI datasets, while addressing error accumulation and anatomical discontinuities.

Key insights

• Incorporating segmentation pseudo-labels (PLs) into the correspondence learning process via OEG-CFN effectively guides the model, reducing error accumulation and improving accuracy in handling anatomical discontinuities.
• Refining initial pseudo-labels with a 3D segmentation model (UNETR++) before feeding them into the OEG-CFN further enhances segmentation performance by leveraging global 3D contextual information.
• The Gradient Enhanced Image Generator, by utilizing second-order derivatives, provides a more robust and precise representation of image edges, leading to more reliable pixel-wise correspondences compared to first-order derivative approaches.

Summary by alphaXiv.