A Combined Computerized Classification System for Whole-slide Neuroblastoma Histology: Model-based Structural Features

Neuroblastoma (NB) is one of the most malignant tumors affecting infants and children. In current clinical practice, NB prognosis and further treatment planning highly relies on histopathological exam- ination of tissue samples. The International Neuroblastoma Pathology Committee has adopted the Shimada classification system, which relies on several morphological characteristics of the tissue such as the degree of Schwannian stromal development and the grade of neuroblastic differen- tiation to categorize the tissue sample as either favorable or unfavorable histology. In this study, we present a combined computer-aided prognosis system that integrates these two diagnosis processes within one analy- sis framework. The proposed system first segments the digitized H&E- stained tissue image into eosinophilic and basophilic structures using the expectation maximization algorithm. For the classification between different tissue subtypes, in addition to conventional co-occurrence tex- ture features, we propose a novel set of structural features that capture higher-level perceptual patterns. We evaluated the developed system over an independent set of 34 whole-slide images and achieved a classification accuracy of 94.1% (32/34).