Reproduce “Handling Supervision Scarcity in Chest X-ray Classification: Long-Tailed and Zero-Shot Learning” 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
Ha-Hieu Pham, Hai-Dang Nguyen, Thanh-Huy Nguyen, Min Xu, Ulas Bagci, Trung-Nghia Le, Huy-Hieu Pham
Researchers from multiple institutions developed methods to address supervision scarcity in Chest X-Ray classification, achieving top performance in the CXR-LT 2026 challenge for both long-tailed multi-label disease recognition and zero-shot identification of unseen pathologies. Their approach combines imbalance-aware training with vision-language models to improve diagnostic accuracy and adaptability in resource-limited scenarios.
Key insights
- Imbalance-aware training, combining distribution-balanced loss and class-aware sampling, effectively addresses long-tailed distributions, improving performance on rare disease findings in multi-label CXR.
- A CXR-specialized vision-language model (WhyXrayCLIP) facilitates robust zero-shot recognition of out-of-distribution findings by aligning visual features with semantic text prompts.
- Ensemble inference, including model ensembling and test-time augmentation, along with post-processing steps like Normal Gating, contributes to overall improved and more stable prediction performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.