ANALYSIS OF MITOSIS DETECTION TECHNIQUES FOR BREAST HISTOPATHOLOGICAL IMAGES

Abstract

The number of mitotic cells is an important metric in the diagnosis of breast cancer. It gives the extent to which the tumor has spread, which has further implications in pre dicting cancer’s aggressiveness. Mitosis counting is a laborious and complex process done manually by a pathologist by analyzing Hematoxylin and Eosin (H&E) stained biopsy slices under a microscope. Detection of mitosis in H&E stained slices is difficult due to insufficient datasets and the similarity between mitotic and non-mitotic cells. The advent of computer aided mitosis detection methods made the whole process far easier by helping in screening, identifying and labeling mitotic cells. Traditional detection methods relied on image pro cessing techniques in which image features are distinguished using handcrafted features. The development and application of neural networks have also been investigated in the mitosis detection process due to their ability to extract features automatically. This project inves tigates mitosis detection as both a classification problem and an object detection problem. The mitosis classification problem employed a multiCNN framework for feature extraction and used machine learning techniques for classification. Tissue level detection of mitosis was also investigated using pre-trained Faster R-CNN by taking raw images as in medical perspective. The experiments were carried out on two publicly available datasets, MITOS ATYPIA-14 dataset and the latest TUPAC16 dataset and the results were compared with other methods in the literature.