Vectorized persistent homology representations for characterizing glandular architecture in histology images

Characterizing glandular architecture in histology images of adenocarcinomas is a fundamental problem in digital pathology, with important implications for computer-assisted diagnosis and grading. In this paper, we present a new set of features for encoding the glandular epithelium architecture based on two recently developed vectorized persistent homology representations called persistence images and persistence landscapes and demonstrate their application to colorectal cancer diagnosis. On the MICCAI2015 Gland Segmentation Challenge Contest dataset with 165 images (85 training, 80 test images), we obtained a benign vs malignant classification accuracy of 85% and 83% using persistence image and persistence landscape based features, respectively.