Recent studies for mitotic figure identification have shown performance comparable to that of human experts; however, the challenge to develop strategies invariant to image variance in different microscope slide scanners still remains. In this paper, we propose a method for domain generalization in mitotic figure detection by considering the scanner as a domain and the characteristic specified for each domain as a style. The method aims to make the mitosis detection network robust to scanner types by using various style images. To expand the style variance, the style of the training image is transferred into arbitrary styles by the proposed style transfer module based on StarGAN. Furthermore, we propose patch selection criteria to resolve the imbalance problem. Our model with the proposed training scheme has obtained satisfactory detection performance on the MIDOG Challenge containing scanners that have not been seen.