Effects of Histopathological Image Pre-processing on Convolutional Neural Networks

Abstract

In this study, classification performance of histopathological images which are processed by pre-processing algorithms using convolutional neural network structure is examined. The images are divided into four different pre-processing classes with their original state and processed with three different techniques. These classes are; original, normal pre-processing, other normal pre-processing and over pre-processing. Histopathological images of these four classes include cancerous and non-cancerous image patches. For these image classes, cancer patch classification is done using the same convolutional neural network structure. In this view, pre-processing effects on the classification success of the convolutional neural network is examined. For the normal pre-processing algorithm, background noise reduction and cell enhancement are applied. For over pre-processing, thresholding and morphological operations are applied in addition to normal preprocessing operations. At the end of the experiments, the most successful classification results are produced with the normal pre-processing algorithms. This is why the meaningful features of the image are left for the CNN structure that automatically learns the feature. The over pre-processing algorithm removes most of these important features from the image.