Observation-Guided Training for Occupational Lung Disease Classification

Separating silicosis and silicotuberculosis in the Silicodata benchmark is notoriously difficult. We improve performance by utilizing radiologist X-ray observations—data previously ignored—to supervise an auxiliary head during training. By discarding this head at inference, the model benefits from expert diagnostic insights without increasing test-time complexity. This technique improves mean class accuracy from $72.3\%$ to $74.6\%$ and sets a new state of the art for AUC across all classes.