Superpixel Segmentation of Breast Cancer Pathology Images Based on Features Extracted from the Autoencoder

Abstract

In order to identify the breast cancer region, it is necessary to discriminate the pathological image of breast cancer pixel by pixel. This is a very huge work for machine learning. Therefore, the preprocessing of superpixel segmentation of breast cancer pathology images is necessary for reducing the number of pixels that need to be discriminated. In this paper: 1. We have trained serveral kinds of autoencoder networks and evaluated their performance in images clustering. 2. In order to enhance the image clustering effect, we adopt the clustering loss which is defined in Deep Embedded Clustering (DEC). 3. In order to enhance the ability of neural networks of extracting features, we added inception-like block, Sequeeze and Excitation (SE) block to the network. 4. We improved the performance of current Simple Linear Iterative Clustering (SLIC) algorithm to achieve superpixel segmentation of high-dimensional features.