Rhabdomyosarcoma Histology Classification using Ensemble of Deep Learning Networks

ABSTRACTA significant number of machine learning methods have been developed to identify major tumor types in histology images, yet much less is known about automatic classification of tumor subtypes. Rhabdomyosarcoma (RMS), the most common type of soft tissue cancer in children, has several subtypes, the most common being Embryonal, Alveolar, and Spindle Cell. Classifying RMS to the right subtype is critical, since subtypes are known to respond to different treatment protocols. Manual classification requires high expertise and is time consuming due to subtle variance in appearance of histopathology images. In this paper, we introduce and compare machine learning based architectures for automatic classification of Rhabdomyosarcoma into the three major subtypes, from whole slide images (WSI). For training purpose, we only know the class assigned to a WSI, having no manual annotations on the image, while most related work on tumor classification requires manual region or nuclei annotations on WSIs. To predict the class of a new WSI we first divide it into tiles, predict the class of each tile, then use thresholding with soft voting to convert tile level predictions to WSI level prediction. We obtain 94.87% WSI tumor subtype classification accuracy on a large and diverse test dataset. We achieve such accurate classification at 5X magnification level of WSIs, departing from related work, that uses 20X or 10X for best results. A direct advantage of our method is that both training and testing can be performed much faster computationally due to the lower image resolution.