Comparison of Deep Learning Preprocessing Algorithms of Nuclei Segmentation on Fluorescence Immunohistology Images of Cancer Cells

Abstract

Immunohistology fluorescence image analysis is an important method for cancer diagnosis. With the widespread application of convolutional neural networks in computer vision, segmentation of images of cancer cells has become an important topic in medical image analysis. Although there are many publications describing the success in application of deep learning models for segmentation of different kind of histology images, the universal model and algorithm of its application is still not developed. Since the histological images of cancer cells are very different, it is usually difficult to get a training set of a large size consisting of images of desired similarity with the studied ones. The image preprocessing consisting in splitting input images in smaller parts and its normalization plays an important role in deep learning especially when the training set is of a limited size. In this study, we compared several approaches to create the training set of a sufficient size while having a very limited number of labeled whole slide immunohistology fluorescence images of cancer cells. In addition, we compared different normalization methods and evaluated their influence on histological image segmentation.