Reproduce “Opportunistic Osteoporosis Diagnosis via Texture-Preserving Self-Supervision, Mixture of Experts and Multi-Task Integration” 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

Jiaxing Huang, Heng Guo, Le Lu, Fan Yang, Minfeng Xu, Ge Yang, Wei Luo

Researchers developed a unified deep learning framework for opportunistic osteoporosis diagnosis from CT scans, addressing challenges in leveraging unlabeled data, mitigating device-specific DXA biases, and integrating clinical knowledge. The framework achieved superior diagnostic accuracy and generalizability across multiple clinical sites and an external cohort, with patient-level F1-scores up to 86.33%.

Key insights

  • Explicitly preserving and utilizing bone texture features through self-supervised learning with a pre-computed texture pool significantly improves the ability to assess osteoporosis from unlabeled CT data.
  • Employing a Mixture of Experts architecture allows a model to dynamically adapt to and regularize against systematic biases introduced by different clinical measurement devices, boosting generalizability across diverse multi-device environments.
  • Integrating quantitative BMD regression and anatomical vertebra position prediction as auxiliary tasks through multi-task learning provides valuable clinical priors, leading to more precise and robust osteoporosis diagnosis.

Summary by alphaXiv.

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