Use of machine learning in bone cancers

Introduction

The medical field, including medical imaging, is amassing data at an increasingly higher rate. Computer-based software, developed to automatically find patterns and infer unknown information from recorded data, is becoming a key component in both medical research and clinical practice. Machine learning (ML)-based software has received special attention in the medical field in the last decade, with a significant number of publications claiming to demonstrate the ability of ML models to achieve or even exceed expert-level diagnosis for multiple diseases, including various cancer types [ ].

Machine Learning is a subfield of artificial intelligence (AI) that refers to a subclass of algorithms which have the ability to improve by learning from data. A ML algorithm, a mathematical model described by a complex mathematical function, learns by analyzing a subset of data of a specific type, called training data. For example, in bone cancer, the training dataset could consist of a number of digitized histology whole slide images (WSIs). The training examples are considered representative of the space of possible data occurrences, and the learned ML model should be able to produce sufficiently accurate classifications or predictions for new, previously unseen data of the same type.

In medical applications, classification usually refers to placing or partitioning the input data (e.g., a WSI) in one of multiple classes, while prediction refers to inferring disease manifestation or patient outcome from the input data (e.g., 5 years survival probability).

The modality of learning from data plays a critical role in the effort involved in building an accurate model. Generally, there are three main approaches: supervised, semi-supervised, and unsupervised. In supervised learning, the algorithm builds a mathematical model from labeled training data. In medical image classification, for example, the input could consist of a WSI, annotated by human experts to reflect regions of viable tumor, necrotic tumor, and other tissue. Through iterative optimization the mathematical model, sometimes called objective function, learns to associate various portions of the WSI with the corresponding class, in this example with viable, necrotic, or other tissue. The most common supervised algorithms include classification and regression [ ]. Classification algorithms are used when the output falls into a small set of classes (e.g., viable tumor, necrotic tumor, other tissue), while regression algorithms are used when the output is a numerical value within a given range. Common supervised ML models include support vector machines (SVMs) and decision trees.

In semi-supervised learning some of the training input samples are missing training labels. Thus, the training set consists of both labeled and unlabeled data. In unsupervised learning the training data is not labeled and the algorithm tries to find structure in the data by identifying commonalities. Clustering and principal component analysis (PCA) are typical examples of unsupervised learning.

Medical applications in general, and bone cancer in particular, have seen the use of supervised models, and our discussion will focus on this class. Currently, there are two main approaches in learning models, one based on traditional ML methods (e.g., SVM, decision trees) and the other one based on deep learning architectures. Deep learning models, which emerged in the last decade, have proven especially successful in medical imaging. Among them, specifically, supervised convolutional neural networks (CNNs) have been the method of choice.

A CNN at a very high level tries to emulate neural connectivity in the human brain. It consists of nodes, emulating neurons, arranged in layers, with nodes in earlier layers connected to nodes in following layers. A CNN has an input and an output layer, with multiple hidden layers in between them, and uses convolution as the main mathematical operation in hidden layers.

In a nutshell, a convolutional layer uses a simple mathematical formula to “convolve” its input and passes the result to the following layer, much like a neuron would respond to a specific stimulus. To reduce the dimension of the data, and through that the number of variables that need to be set to obtain a working model, the convolution layers are intermixed with pooling layers by combining the outputs of node clusters at one layer into a single node in the next layer. It is beyond the scope of this presentation to get into more details on the structure and inner workings of CNNs. Such details can be found in various publications, including Refs. [ , ]. When talking about CNNs we will interchangeably use the words model and architecture, although architecture specifically might refer to the design of the hidden layers, including the number of convolutional and pooling layers, and how they are spaced.

Next, we present the most recent ML models developed for bone sarcoma. They include both traditional ML models and CNN models.