Impact of cancer histopathological image preprocessing on convolutional neural network performance

Abstract

Histopathological image analysis has been the golden standard for cancer diagnosing and grading. With the recent advances in deep learning techniques for digital image analysis, convolutional neural networks (ConvNets) have been successfully used to aid pathologists with the time-consuming and laborious examinations. To produce correct diagnoses, the pathologists visually examine cells' shape and organization along with overall tissue structure and morphology by observing the microscopic slides through different magnification levels. Based on this, current deep learning approaches are developed by extracting small patches from the whole slides in very distinct manners, altering the size of the patches, the magnification level and, the normalization method

factors that can change the performance of the ConvNet in terms of accuracy and efficiency. Therefore, in this study, we train a total of twelve ConvNet models to perform the task of 3-class classification of stomach cancer histopathological images, varying the size of the patches, the magnification level, and the normalization method, to analyze the effect that each factor has on the convolutional neural network, assessing the performance via precision and recall. With this study, we provide a rationale for why these factors would affect a model's performance in relation to data representation, as well as a guideline for histopathological image preprocessing methods.