Mitotic count is an important feature for breast cancer diagnosis but their striking resemblance with non-mitoticgures, no proper shape, and scarce number makes their detection a laborious and challenging task. This thesis presents an integrated system for automatic scoring of breast cancer Whole Slide Images. To deal with the imbalance between mitotic and non-mitoticgures a two-phase learning strategy is proposed, where therst phase informatively undersamples the majority class so that the second phase can concentrate more on hard examples. To harness the rich features extraction and mapping capabilities of Deep Neural Networks in case of small dataset, Transfer Learning based segmentation and classi cation is proposed. Finally, to tackle the large sized Whole Slide Images an e ective and e cient method is proposed for region of interest selection and scoring the slides. The integrated system comprising of region of interest selection, mitosis detection, and slide scoring achieved state-of-the-art results with a Kappa score of 0.5823 on a publicly available dataset and constituted a major step towards clinical application of Computer Assisted Diagnosis for the good of humanity.