Convolution kernel size effect on convolutional neural network in histopathological image processing applications

Abstract

In this study, the change in the classification success of the convolutional neural network (CNN) is investigated when the dimensions of the convolution window are altered. For this purpose, four different linear convolution neural network architectures are constructed. The first architecture includes 4 convolution layers with 3×3 convolution window dimensions. The second architecture includes 4 convolution layers with 5×5 convolution window dimensions. The third architecture includes 4 convolution layers with 7×7 convolution window dimensions. The fourth architecture includes 4 convolution layers with 9×9 convolution window dimensions. A dataset consisting of histopathological image patches is used to test the CNN architects that are created. 2000 training images and 250 validation images on dataset are applied to all architectures with the same order, in order to fair assessment. In conclusion, the effect of convolution dimensions on classification of histopathological images by deep learning methods is determined. The test results of four different linear convolutional neural network architectures are evaluated using sensitivity, specificity and accuracy parameters. © 2018 IEEE.