Hybrid Feature Set based Mitotic Detection in Breast Histopathology Images

• Main Authors: Deepak Naik, M V, Adarsh, S, Vijayakumar, Sreekumar, Nambiar, Abhay P, Nair, Lekha S
• Format: Conference Proceeding
• Language: English
• Subjects: Breast cancer; Breast Histopathology; correlation-based feature selection; Feature extraction; grey level co-occurrence matrix; Histopathology; Image edge detection; Image segmentation; Shape; Training; whole slide imaging
• ISSN: 2767-7788
• DOI: 10.1109/ICICT54344.2022.9850552

Breast cancer is a very serious threat to women and requires faster diagnosis methods to decrease the mortality rates. Automatic detection of mitosis is very difficult when compared to other pattern detection since mitotic cells are irregularly shaped objects, therefore there is no simple way of extracting features of a mitotic cell. This paper proposes a novel method with which a dataset comprising of shape and texture features are created from the image patches extracted from whole-slide histopathology images (MITOS-ATYPIA-14). The extracted image patches are converted to blue ratio images and hysteresis thresholding is applied for better edge detection. The ROI is extracted and a set of 10 shape features and 34 texture features are then extracted from each segmented image to form a hybrid feature dataset. Correlation-based feature selection is used to obtain the dominant features which can classify the data points as mitotic and non-mitotic. The dataset is then used to train seven models and classification is performed on unannotated image patches and the best model is selected based on the accuracy.