Computer-Based Classification of Nuclei in Gliomas.
Jun Kong, Lee AD Cooper, Candace S Chisolm, Fusheng Wang, Carlos S Moreno, Tahsin Kurc, Daniel J Brat, Joel H Saltz. Emory University, Atlanta, GA
Background: We are conducting computer-based morphological studies of diffuse gliomas to identify morphological correlates of patient outcome and genomic characterizations. Previous work using The Cancer Genome Atlas glioblastoma data demonstrated morphological differences between tumors of specific transcriptional classes when averaged nuclear features within a tumor are examined. In this study we subclassified different types of cell nuclei encountered in gliomas using machine-based algorithms in an effort to increase the resolution of correlative morphological investigations.
Design: We used a set of 16 digitized H&E-stained images to collect neuropathologists' classifications of 2240 nuclei as neoplastic astrocytes, neoplastic oligodendrocytes, reactive endothelial cells, reactive astrocytes, and hematoxylin-stained debris. The slides were chosen from a spectrum of gliomas, including astrocytomas, oligodendrogliomas, oligoastrocytomas, and glioblastomas (which contained 0+,1+,2+ oligo component). Nuclei in these images were segmented using computer algorithms. A set of features describing nuclear shape, texture and cytoplasmic staining was calculated to describe each nucleus. Features derived from the neuropathologist-classified nuclei were used to train a computer classification system. A feature selection algorithm first identified a subset of features that best captures class differences. Selected features were then used to train a quadratic discriminant classifier. A five-fold cross validation was used to calculate the average classification accuracy as the ratio of the number of correctly classified nuclei to that of total nuclei. The bottom 10% of ambiguous nuclei as ranked by classification confidence score were discarded from each fold.
Results: We found that the overall computer-based classification accuracy was 92%. The classifier achieves promising accuracy for neoplastic oligodendrocytes (95.6%), neoplastic astrocytes (91.9%), reactive endothelial cells (89.0%), and debris (89.9%). Reactive astrocytes (85.6%) are the most difficult to classify, with 5.3% being mislabeled as neoplastic astrocytes and 6.2% being misclassified as reactive endothelial cells.
Conclusions: These results suggest that computational morphometry can achieve reasonable classification rates for multiple nuclear classes commonly encountered in the gliomas. We are currently extending the classification scheme to include microglia, mitotic figures, and the normal nuclear types that are critical for studies on low-grade gliomas.
Tuesday, March 1, 2011 1:00 PM
Poster Session IV # 192, Tuesday Afternoon