Towards Automatic Identification and Delineation of Tissues and Pathologies in H&E Stained Images

We propose a framework and methodology for the automated identification and delineation of tissues and their pathologies in hematoxylin and eosin (H&E) stained images. Histopathology is vital to medicine and research as it enables quantitative and qualitative analysis of tissue samples, stained and visualized via microscopes; the most routine and cost-effective of these stains is H&E. In clinical diagnostic surgical pathology, the pathologist interprets H&E-stained tissue slides by determining whether a given sample represents normal or abnormal tissue for the given anatomical location. Although pathologists accurately and consistently identify and delineate such tissues and their pathologies, this is a time-consuming and expensive task; thus the need for automated algorithms for improved throughput and robustness.

We develop such an algorithm that uses local histograms and occlusion models as a mathematical framework for pixel-level classification. We also develop an expert-guided feature set called the histopathology vocabulary that mimics the visual process used by pathologists. To expand applicability, we achieve simultaneous identification and delineation by performing pixel-level classification.

Experimental results on both a clinical application (active colitis) and a research one (tissue development in teratoma tumors) validate the discriminative power of our approach. We also present comparisons to popular, though general, feature types to demonstrate the power of our expert-guided feature set. Our framework and methodology demonstrates great promise towards the creation of a framework and methodology for the automated identification and delineation of tissues and their pathologies in H&E-stained images.