Machine Learning for Rhabdomyosarcoma Whole Slide Images Sub-type Classification

The most frequent malignant soft tissue tumor in children is Rhabdomyosarcoma (RMS). RMS has several subtypes that differentiate treatment and patient outcomes. Because of variations in the appearance of histopathology images, manual subtype classification requires a high level of expertise and is time-consuming. While several machine-learning techniques have been developed to classify the most common tumor types in histology images, more must be understood about the automatic classification of tumor subtypes. Moreover, existing techniques for classifying whole slide image (WSI) histopathological subtypes rely on small, randomly selected image tiles or on representative tiles chosen by specialists from the considerably larger WSIs. These methods do not draw knowledge from the whole tissue region captured by a WSI, possibly missing significant tumor signatures. They also fail to account for the spatial distribution of patterns that could play a role in reliable subtype classification. This paper proposes a novel methodology that combines different methods to extract features from a whole slide image and generates a whole slide feature map (WSFM). This map and additional clinical features are then used to train various machine-learning models. We obtain 91.84% WSI tumor subtype classification accuracy on a diverse dataset. A direct advantage of our methodology is that it does not require any WSI-level annotation by pathology experts. Training and testing can be performed much faster computationally using simple machine learning algorithms instead of complex deep learning architectures.