The aim was to estimate the incidence of Mycobacterium tuberculosis (Mtb) infection in health-care workers (HCWs) in Japan. We repeated cross-sectional surveys of HCWs with QuantiFERON®-TB Gold (QFT-G) in 2003, 2005 and 2007 at a hospital with tuberculosis (TB) wards, and 311 HCWs who underwent QFT-G testing two or three times were included in the study. Five HCWs (1·8%) converted from negative to positive. Incidence of new TB infection was estimated to be 0·6/100 person-years by the CDC's definition. Thirteen positive persons (41%) reverted from positive to negative. Multivariable logistic regression analysis identified a significant association between QFT-G conversion and working in TB wards. The IFN-γ levels of all but two subjects with reverting or converting QFT-G results were close to the test's cut-off. The incidence of Mtb infection in HCWs at our hospital was higher than that estimated for the general population in Japan. Criteria for defining QFT-G conversion and reversion need further investigation considering the high proportion of reversion, as the incidence of infection would have changed if we had applied other definitions.