Histopathological cancer detection: a comparison study of different convolutional neural networks

The histopathological method is important for cancer diagnosis. The standard histopathological test today is extremely time-consuming, expensive, and labor-intensitive. However, as image digitalization technology became ubiquitous today, automatic analysis using computer-aided diagnosis tools provides a possible alternative. Using deep learning, especially for convolutional neural networks, it is possible to obtain results that have comparable performance to pathologists with a higher diagnosis speed. In this paper, seven different CNN architectures were trained on the Histopathological Cancer Detection dataset. The dataset contains 20000 images extracted from histopathological scans of lymph node sections. The results showed that there is a correlation between the performance and the architecture design’s appropriateness on a specific dataset. DenseNet achieved the best result on this dataset with an accuracy of 0.9182 without data augmentation, and 0.9268 with data augmentation. One possible reason for the superior performance of DenseNet is that DenseNet mimics the diagnosis process of pathologists by integrating image features with multiple different scales. The experiment results suggest that DenseNet can serve as an effective tool to speed up the diagnosis and ease the workload of pathologists.