Machine learning

ROSA VERHOEVEN: AI is a very broad term that encompasses many methodologies used to perform a task that would normally require human intelligence. This could already be a very simple system using hard-coded, expert knowledge to support a decision. A more advanced subset of these methodologies fall under the category of machine learning, in which the machine is trained on example data so it can learn to solve a particular problem autonomously. In this video, we will focus on how this machine learning works. Machine learning can be split up into three different categories– supervised learning, unsupervised learning, and reinforcement learning. Let’s discuss each of these categories.

Supervised learning is the most common type of machine learning, which can only be used in case of data in which the machine will be trained is accompanied by a particular label. Suppose you have a collection of CT slices that contain either a benign or a malignant tumour. The label of the individual images is then either benign or malignant. These examples are passed on to the machine-learning model, which trains on matching these slices to their labels. If unlabeled CT slices are then passed on to the model, it will try to predict their label based on what it has learned from the example data. This is called classification.

An alternative to classification is regression, in which the labels are numerical instead of categorical. An example of regression is when the algorithm predicts the likelihood of malignancy. Finally, labels can also consist of so-called annotations. In the example, this could mean a contour in the image drawn by an expert that indicates the lesion boundary. In the case of unsupervised learning, there are no labels provided with the training data. Instead, these types of algorithms try to group the data based on natural associations. The machine looks at the example data and tries to figure out which features are most important to be able to separate the data.

In the case of the previous example, the algorithm would try to group malignant and benign CT slices based on their common visual elements, without explicitly labelling these groups. The last type of machine learning is called reinforcement learning. This type of learning utilises a trial-and-error approach to accomplish a particular task, usually in a complex environment. This is executed with the use of an intelligent agent, which is basically anything that can perceive its environment, take actions, and improve its performance using knowledge and experiences. An example of an intelligent agent is a robot. The agent is placed into a particular environment.

And by receiving rewards and punishments when the agent is in a particular state based on the actions it has taken, it learns how to match those actions and situations with the reward. One application in a medical context is the training of algorithms to learn how to run to be implemented into prosthetics. Another example can be applied to our previous situation of tumours in the CT slices. That is, in that situation, reinforcement learning could mean a continuously-learning system that retrains based on the acceptance or rejection of its predictions. For each of the three types of machine learning, there are various algorithms that can be used.

For example, supervised learning can be done with the use of an algorithm called a decision tree, which actually resembles an upside-down tree. At each split of the decision tree, also called a node, a feature is used to split up the data. Is the value of the data lower than x? The data goes left. Is it higher than x? The data goes right. Then the next node is encountered, in which a new feature is chosen to split up the data. This continues until a leaf– or the end of a branch– is reached, which tells you to which category or label the data belongs. For unsupervised learning, one type of algorithm that can be used is called K means clustering.

This type of algorithm attempts to cluster the data together into a particular number of clusters, K, based on the spreading of the features. However, for all three types of machine learning, it’s also possible to use artificial neural networks instead of these specific algorithms mentioned before. If these networks contain a lot of neuron layers, then you’re talking about deep learning. Artificial neural networks are based on the concept of biological neural networks, and the way in which they work does resemble human intelligence and learning very closely. Applications using artificial neural networks have shown to be very promising and are currently the main point of focus for AI in healthcare.