Comparison of deep learning architectures for H&E histopathology images

Deep learning has achieved outstanding performance in many fields such as image classification and target recognition. Recently multiple research efforts are focusing on deep learning to medical image processing. While it is common in image processing to apply transfer learning for problems with small sample sizes, the statistics of histopathological stains are known to be very different from the photographic RGB images in common deep learning imaging tasks such as Imagenet and MIT Places. This paper evaluates the performance of fine-tuned models on Haematoxylin and Eosin(H&E) histopathology stain data. Furthermore, to analyze the performance of different deep learning architectures on these domains, we compare three convolutional neural network(CNN) architectures in various settings. Finally, the impact of the size of the context of training samples is evaluated. We use the BreaKHis dataset consisting of H&E stained microscopical scans of breast cancer tissue [1]. Our results show that fine-tuned architectures perform favorably over neural networks that are trained from scratch in terms of accuracy and patient rate.