Deep Learning Models Differentiate Tumor Grades from H&E Stained Histology Sections

Aberration in tissue architecture is an essential index for cancer diagnosis and tumor grading. Therefore, extracting features of aberrant phenotypes and classification of the histology tissue can provide a model for computer-aided pathology (CAP). As a case study, we investigate the application of convolutional neural networks (CNN)s for tumor grading and decomposing tumor architecture from hematoxylin and eosin (H&E) stained histology sections of kidney. The former and latter contribute to CAP and the role of the tumor architecture on the outcome (e.g., survival), respectively. A training set is constructed and sample images are classified into six categories of normal, fat, blood, stroma, low-grade granular tumor, and high-grade clear cell carcinoma. We have compared the performances of a deep versus shallow networks, and shown that the deeper model outperforms the shallow network.