A comparative analysis of sensitivity of convolutional neural networks for histopathology image classification in breast cancer

Abstract

The Convolutional neural networks (CNN) have been shown to be able to learn the relevant visual features for different computer vision tasks from large amounts of annotated data. Hence, the performance of CNNs can vary depending on the training data set and associated model architecture. This article presents a comparative analysis of the robustness and sensitivity of different CNN architectures to classify invasive breast cancer tissue. Our experiment involved a comparison of six CNN architectures with different depths (number of layers), specifically trained to detect invasive breast cancer from digitized pathology images. Additionally, the pre-trained model VGG 16 (trained to classify natural images) was added as the seventh architecture. Each of the models was trained with two different data sets: a cohort of 239 breast cancer slide images from the Hospital of the University of Pennsylvania (HUP), and another with 172 digitized breast cancer images from the Cancer Genome Atlas (TCGA). In addition, in each case the training was validated with 40 breast cancer slide images from the New Jersey Cancer Institute (CINJ). The last layer of the VGG 16 model was modified to allow classification of the binary problem (presence or absence of invasive ductal carcinoma). The experimental results show a performance of greater than 93% in terms of AUC (Area Under the ROC Curve) for the CNNs trained specifically with cases of invasive breast cancer from the TCGA. However, we also note that VGG-CNN-16 achieves an AUC of 92.43% and 86.87% respectively, despite the fact that it was trained for a different domain.