Molecular Classification of Ependymomas Using Histological Images and Deep Neural Networks

Ependymomas represent a rare type of tumor in the central nervous system that affects both
children and adults. For these tumors, strong differences in quality of diagnostic healthcare exist
between medical centers in Germany. Thus, molecular analyses (e.g., DNA methylation profiling)
are increasingly used to validate the pathological diagnoses from traditional examination of
histological images. However, these molecular analyses are expensive and are not readily available
worldwide. Thus, we employ deep neural networks to predict the molecular properties of
ependymomas from large-scale, histological image data and aim to support neuropathologists in the
integrated diagnosis of this challenging tumor entity.

Key findings:

• Using histological image data, the molecular (DNA methylation) type can be accurately predicted for ependymomas from all major anatomical compartments
• Color normalization, color augmentation and the choice of multiple-instance pooling operation are major factors that determine domain-adaptation to other imaging facilities
• Simple code optimization steps facilitate highly efficient data processing on the supercomputer HSUper