WALDO Framework Enhances Zero-Shot Anomaly Localisation in Medical Imaging Using Vision-Language Models

Research presents WALDO, a training-free framework for zero-shot anomaly localisation in medical imaging, leveraging vision-language models (VLMs) to improve rare pathology detection. The framework reformulates anomaly detection as a comparative inference problem, utilizing entropy-weighted Sliced Wasserstein distances for anatomically-aware reference selection, Goldilocks zone sampling for optimal reference similarity, and self-consistency aggregation through weighted non-maximum suppression. Theoretical analysis indicates that moderate similarity references minimize bias-variance trade-offs in visual reasoning. Evaluated on the NOVA brain MRI benchmark, WALDO with Qwen2.5-VL-72B achieved 43.5% mAP@30, a 19% relative improvement over zero-shot baselines, with statistical significance confirmed by paired McNemar tests ($p<0.01$). Source code is available on GitHub.