Reproduce “Orthogonal Annotation Benefits Barely-supervised Medical Image Segmentation” on OpenResearch

About this paper

Heng Cai, Shumeng Li, Lei Qi, Qian Yu, Yinghuan Shi, Yang Gao

An annotation method and co-training framework from Nanjing University enables 3D medical image segmentation with only two orthogonal slices per volume, drastically reducing labeling effort. This approach achieved Dice scores up to 89.24% on the LiTS dataset and 86.93% on KiTS19, often outperforming state-of-the-art methods reliant on substantially more labeled data.

Key insights

• Orthogonal slices provide more diverse and complementary information (lower Hilbert-Schmidt Independence Criterion scores) compared to parallel slices, leading to more consistent feature representations for the same class.
• Gradually shifting supervision emphasis from dense (noisy) pseudo-labels to sparse (precise) ground truth annotations, combined with increasing cross-supervision, is crucial for stable and high-performance training with minimal annotations.
• Uncertainty-guided pseudo-label selection in cross-supervision effectively mitigates error propagation between networks, improving learning from unlabeled data.

Summary by alphaXiv.