Multi-view Learning and Deep Learning for Microscopic Neuroblastoma Pathology Image Diagnosis

Automated pathology image diagnosis is one of the most crucial research in the computer-aided medical field, and many studies on the recognition of various cancers are currently actively conducted. However, neuroblastoma, the most common extracranial solid tumor of childhood, has not got enough attention in the computer-aided diagnosis research. Accurate diagnosis of this cancer requires professional pathologists with sufficient experience, which makes lack of experts lead to misdiagnosis. In this paper, we apply multi-view and single-view maximum entropy discrimination, with traditional image representations and deep neural network representations respectively. The diagnosis is performed in three neuroblastoma subtypes, undifferentiated subtype ( UD), poorly differentiated subtype ( PD), differentiating subtype ( D), and the normal type un-neoplasm tissues ( UN). The best classification performance ( 94.25%), which far exceeds the diagnosis accuracy ( 56.5%) of a senior resident in the corresponding field, demonstrates the potential of neural network representations in analyzing microscopic pathology images of neuroblastoma tumors.