Grading of tumor malignancy in breast cancer should contribute essential information both for the prospective outcome of the individual patient as well as for TNM staging. In a series of 104 breast cancer patients we tested the prognostic validity and reproducibility of mitotic figure counting compared with TNM staging, Bloom and Richardson grading, DNA single cell cytometry, and morphometry. Four-micrometer thick hematoxylin-eosin-stained routine slides were investigated. Mitotic figures were counted in representative areas of the tumor in 10 159-microns 2-sized high power fields (HPFs) at a 400X magnification; the median value was seven and the threshold for the 25th percentile was three. This value should replace the common but prognostically invalid threshold of 10. Univariate survival analysis showed that mitotic figure counting allows the identification of three groups of patients (< or = 3, 4 to 20, > 20 mitoses per HPF) with significantly different survival probabilities (P < .0001; P = .0178). Depending on the number of mitotic figures, length of survival was significantly different within the group of T1N0 tumors (P = .0082) and the group of T1N1 or T2N0 tumors (P = .0251). In a Cox stepwise regression model mitotic frequency counting added prognostic information to tumor size and was of higher prognostic significance than lymph node status, DNA ploidy, or mean nuclear area. The 95% confidence limit for interobserver reproducibility, tested in 20 cases, was plus/minus 8 mitoses. After quartilization an agreement of 75% was observed.