Abstract:
Mitotic count is an important feature for breast cancer diagnosis but their striking
resemblance with non-mitotic gures, 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-mitotic gures a two-phase learning strategy
is proposed, where the rst 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.
