Keywords: Digital Pathology, Deep Learning, Convolution Neural Network
Digital pathology images of tumor specimens have been scanned, represented and archived for every patient for all oncology trials. These images present unique and detailed geo-spatial information of tumor cell micro-environments, and have been playing key roles in pathological analysis and diagnostic medicine for over a century. However, the extra-high-dimension and high-correlation natures of those images impose many technical hurdles for conventional human interpretation. The development of Deep Learning, especially convolution neural networks (CNNs) for analyzing and classifying image patterns have led to tremendous success in this area. In this study, we extensively explore the practical aspects of CNNs implementations in digital pathology: 1) if and what CNN architectures are suitable for different stages of the digital pathology image analysis; 2) what are the optimal input image sizes for CNN modeling; And 3) how can we tune the hyper parameters of the CNN models. We demo our optimized approaches in addressing these pressing technical issues in digital pathology analysis on the public TCGA-SKCM H&E whole slide image data set.