The accurate and rapid assessment of tumor margins during breast cancer resection using optical coherence tomography (OCT) has the potential to reduce patient risk. However, it is difficult to subjectively distinguish cancer from normal fibroglandular stromal tissues in OCT images, and an objective measure is needed. In this initial study, we investigate the potential of a one-dimensional fractal box-counting method for cancer classification in OCT. We computed the fractal dimension, a measure of the self-similarity of an object, along the depth axis of 44 ultrahigh-resolution OCT images of human breast tissues obtained from 4 cancer patients. Correlative histology was employed to identify distinct regions of adipose, stroma, and cancer in the OCT images. We report that the fractal dimension of stroma is significantly higher than that of cancer (P < 10(-5), t-test). Furthermore, by adjusting the cutoff values of fractal dimension between cancer, stroma, and adipose tissues, sensitivities and specificities of either 82.4% and 88.9%, or 88.2% and 81.5%, are obtained, respectively, for cancer classification. The use of fractal analysis with OCT could potentially provide automated identification of tumor margins during breast-sparing surgery.