An Ensemble Algorithm Framework for Automated Stereology of Cervical Cancer

Stereological procedures to quantify mean nuclear volume are commonly used to differentiate cancerous from normal tissue. Automatic quantification of these parameters requires segmentation, which is complicated by the variability in tissue staining and nuclei size. One solution to deal with such alterations in a robust fashion is to use an ensemble of segmentation methods. The goal of this work is to demonstrate the use of an ensemble of simple segmentors in a novel way to improve the performance achieved by the individual segmentors. The contributions of this paper are three fold: applying an ensemble on the blob level in addition to the image level, utilizing the image level ensemble to accept or reject input images based on their segmentation quality and finally applying the ensembles for discriminating cancer and normal classes. Hematoxylin and eosin (H&E) stained sections from archival tissues from the normal cervix and cervical cancer have been used as the dataset. The results presented here show that both levels of ensembles enable clear class separability as compared to the individual segmentors, and thus demonstrate the effectiveness of the proposed ensemble framework.