Multi-pixels Classification for nuclei segmentation in digital pathology based on deep machine learning

Abstract

The problem this paper solves is identifying the accurate boundary for nuclei in breast cancer digital pathology images. By doing the nuclei segmentation, the output image can be a reference for the doctors to make better diagnosis for breast cancer patients. Using convolutional neural network to solve this problem which is a popular deep learning method for image classification. By generating labelled patches from the original dataset for training and using a sliding window to cut out patches from testing image for classifying pixels in the testing image, the image level classification can be transferred to pixel level classification. A 2x2 receptive field can be classified to 16 classes, and each pixel can be calculated 4 time, which increase the pixel level TPR (true positive rate) from 0.868 to 0.909. And the boundaries are smother than traditional method that the receptive field is 1x1.