Computer-assisted Prognosis of Neuroblastoma: Grading of Differentiation
Hiroyuki Shimada MD; Childrens Hospital Los Angeles, The University of Southern California; Olcay Sertel MSc; Department of Biomedical Informatics, The Ohio State University; Metin Gurcan PhD; Department of Biomedical Informatics, The Ohio State University; Kim Boyer PhD; Department of Electrical and Computer Engineering, The Ohio State University; Joel Saltz MD; Department of Biomedical Informatics, The Ohio State University; Jun Kong MA; Department of Biomedical Informatics, The Ohio State University;
Content:
Neuroblastoma is an embryonal tumor originated from the sympathetic nervous system. In compliance with the International Neuroblastoma Prognosis Classification System developed by Shimada et al., neuroblastoma can be categorized into undifferentiated, poorly differentiated and differentiated groups in terms grade of differentiation. In current clinical practice, neuroblastoma classification is carried out by highly trained pathologists with visual examinations of pathological slides, a process that is time consuming and prone to inter- and intra-reader variability. Thereby, we are motivated to develop a computer-aided grading system that helps pathologists in their prognosis.
Technology:
A multi-resolution and multi-classifier computerized system is developed to automate the classification process. The inputs to the system are digital images of tumor samples that are sliced at 5 ?m in thickness, stained using haematoxylin and eosin, digitized with the ScanScope T2 digitizer (Aperio, San Diego, CA) at 40 x magnification and then compressed at a 1:40 compression ratio. The basic components of the system include image decomposition, image segmentation, feature construction, feature selection, feature extraction, classification, classifier combination and classification evaluation. The output of the algorithm assigns to the given input image a label representing its grading type.
Design:
This developed system uses a coarse-to-fine evaluation strategy that mimics those followed by pathologists. The designed algorithm begins the classification process by producing image representations of different resolution levels and initially works on lower-resolution images, which correspond to lower optical magnification levels of microscopes. In areas where additional image details are needed, the computerized system is dynamically directed to work on higher-resolution image representations. The decision is made automatically by voting and weighting prior classifier accuracies of multiple classifiers integrated into the system. Additionally, we proposed a novel segmentation approach that encapsulates the Fisher-Rao criterion into the generic Expectation-Maximization algorithm.
Results:
The classification accuracy was tested by comparing the computer decisions with those of the manual classification. With our testing studies, consisting of 14 undifferentiated, 12 poorly-differentiated and 18 differentiating tumor slides, the developed system produced the best classification accuracy of 84.09% (36/44), with class conditional accuracies of 13/14, 9/12 and 15/18 respectively. Additionally, the combination of multi-resolution schema with the automated feature selection process resulted in a 34% savings in computation time when compared to a previously developed system.
Conclusion:
The developed system shows promising classification accuracies for computer-aided pathological assessment of grade of neuroblastic differentiation on a whole-slide dataset with significant savings in computational costs.
