Construction and mine dust: Michigan AI identifies first lung lesions at 91%, compared to 77% for human readers

Trained on X-rays of U.S. workers, Michigan State's tool supports the 200 certified specialists still practicing across the country.

Researchers at Michigan State University have put together a dataset of chest X-rays of American workers — the first built specifically from this population — to train a program to detect early lung lesions. Published in Occupational and Environmental Medicine, the study reports a 91% accuracy in identifying the very first fibrosis micro-opacities, where human readers average 77%.

The public health problem is massive: more than 2.2 million American employees inhale rock, sand or coal dust every day in mines, construction or the manufacture of worktops. The trapped particles end up causing irreversible pathologies, silicosis or black lung. However, early diagnosis is based on the reading of X-rays by specially certified doctors, the "B readers": there are only 200 of them left in practice in the United States.

The tool is presented as an assistant rather than a substitute. It intervenes on four screening tasks, superimposes a colored map on the image to designate suspicious areas, and has succeeded in removing about half of the healthy images from the review queue. The stated objective is less to replace the expert eye than to concentrate it on doubtful cases, and to speed up the issuance of documents that condition workers' compensation.