Reproduce “Dual-Scale Volume Priors with Wasserstein-Based Consistency for Semi-Supervised Medical Image Segmentation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Junying Meng, Gangxuan Zhou, Jun Liu, Weihong Guo

This paper introduces VP-Net, a semi-supervised medical image segmentation framework that integrates dual-scale volume priors and Threshold Dynamics spatial regularization. The method leverages a learnable regression network for adaptive image-scale volume prediction and a novel dataset-scale volume distribution prior enforced via Wasserstein distance, achieving improved accuracy and robustness on various medical imaging datasets.

Key insights

  • Directly embedding variational model-derived priors (spatial regularization and volume preservation) into the deep network's final layer significantly enhances performance and interpretability in semi-supervised settings.
  • Leveraging dual-scale volume priors—a strong image-scale prior through adaptive prediction and a weak dataset-scale distribution prior via Wasserstein distance—provides comprehensive and effective regularization for unlabeled data.
  • Adaptive, learnable prediction of image-specific volumes is crucial for effective regularization, outperforming fixed, empirical average volumes.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.