A Comprehensive Framework for Classification of Nuclei in Digital Microscopy Imaging: An Application to Diffuse Gliomas

• Type of Material: Article
• Authors: Jun Kong, Emory University; Lee Cooper, Emory University; Fusheng Wang, Emory University; Candace Chisolm, Emory University; Carlos Moreno, Emory University; Tahsin Kurc, Emory University; Patrick Widener, Emory University; Daniel Brat, Emory University; Joel Saltz, Emory University
• Language: English
• Date: 2011-01-01
• Publisher: IEEE
• Publication Version: Author Accepted Manuscript (After Peer Review)
• Copyright Statement: © 2011 IEEE.
• Final Published Version (URL): http://dx.doi.org/10.1109/ISBI.2011.5872833
• Title of Journal or Parent Work: 2011 8TH IEEE International Symposium on Biomedical Imaging: From Nano to Macro
• Start Page: 2128
• End Page: 2131
• Grant/Funding Information: This research is supported in part by Grant Number R01LM009239 from the NLM, by NCI Contract No. N01-CO-12400 and 94995NBS23 and HHSN261200800001E, and by NSF CNS 0615155, and 79077CBS10.
• Keywords: Engineering; microscopy image analysis; Technology; Science & Technology; diffuse glioma; metadata model; Radiology, Nuclear Medicine & Medical Imaging; Life Sciences & Biomedicine; feature selection; Nuclei classification; Engineering, Biomedical
• Research Categories: Engineering, Biomedical

Abstract

In this paper, we present a comprehensive framework to support classification of nuclei in digital microscopy images of diffuse gliomas. This system integrates multiple modules designed for convenient human annotations, standard-based data management, efficient data query and analysis. In our study, 2770 nuclei of six types are annotated by neuropathologists from 29 whole-slide images of glioma biopsies. After machine-based nuclei segmentation for whole-slide images, a set of features describing nuclear shape, texture and cytoplasmic staining is calculated to describe each nucleus. These features along with nuclear boundaries are represented by a standardized data model and saved in the spatial relational database in our framework. Features derived from nuclei classified by neuropathologists are retrieved from the database through efficient spatial queries and used to train distinct classifiers. The best average classification accuracy is 87.43% for 100 independent five-fold cross validations. This suggests that the derived nuclear and cytoplasmic features can achieve promising classification results for six nuclear classes commonly presented in gliomas. Our framework is generic, and can be easily adapted for other related applications.