Mitosis Detection in Multispectral Histopathological Images with Deep Learning

In this study, segmentation of cellular structures in the multispectral histopathological images and possibility of the discrimination within normal and mitotic cells have been investigated. In histopathological images, it is very challenging task to extract the mitotic cells from the histopathological image. In the first stage of the study, `discriminative images' are obtained using linear discriminant analysis. The discriminative images found are used to screen mitotic cell candidates and train them with deep learning networks. Since mitotic cells are usually dark pixels, the discriminating images are first filtered to obtain dark areas. Then, two-clustered k-means algorithm was used to differentiate background and mitotic candidates. With the help of convolutional neural networks, it is aimed to find the mitotic and non-mitotic areas with the classification approach. In experimental studies, ICPR-2012 dataset is used in both training and prediction stages and the training of deep artificial neural network architecture is carried out with samples of mitotic and non-mitotic regions. As a result, the F-measure for the data set is found as 0.760, the recall is 0.723 and the sensitivity is 0.802.