Use of machine learning in bone cancers

Machine learning models that assist physicians in the decision-making process have been developed and implemented in the last few years. Generally, such models are used to classify digital data into clinically relevant subclasses, such as diagnostic tumor types (osteosarcoma, Ewing sarcoma, rhabdomyosarcoma, etc.) or histological tissue subclass (viable, necrotic, or healthy tissue) from digital histology data. Practically a model, which encodes a complex mathematical function, is developed by learning a set of features from training datasets of digital images (e.g., radiology and/or histology). It then uses the learned features to infer information from previously unseen datasets. Experience in osteosarcoma to date emphasizes the use of two major model types, based on the underlying model architecture. The first class contains traditional machine learning models, implementing supervised learning to extract a set of well-defined features from the input data to produce the output classification. The second class is based on emerging deep learning architectures, such as convolutional neural networks, that implement a supervised learning approach to an annotated training dataset and uses learned features to classify future input data. However, a key particularity of deep learning models is that the features learned are abstract to the user, numerical encodings that cannot be directly associated to human expert defined features of specific classes. We review recent progress on developing computer-based classifiers for bone cancer that include both traditional machine learning models and convolutional neural network architectures.