The specifics of how a machine learning algorithm is trained to recognize certain features and thereby become able to make accurate predictions on new examples varies depending on the type of data being used and the algorithm architecture. Four of the most commonly used learning processes are:
-
supervised learning
- most commonly applied in radiology
- typically a set of images (training set) is provided along with a known feature or diagnosis (ground truth); although supervised learning is also used on radiology reports to develop natural language processing algorithms
- machine learning algorithms makes predictions based on response values usually towards the goal of classification but supervised learning can also be used for regression or other models
- because the input data (training set) and the response values are already known (ground truth) the algorithm can make iterations until it reaches an agreed-upon result
-
unsupervised learning
- less common in radiology
- the algorithm is only fed input data with no known ground truth
- algorithm architectures can include clustering, neural network models, or other models
- often used to identify trends or patterns within a data
-
reinforcement learning
- applied in the context of an agent inspecting its surroundings and performing actions in order to receive delayed rewards
- even less commonly used in radiology but recently used in image analysis tasks for high-resolution images in order to focus on inspecting the relevant areas
-
evolutionary algorithms
- a population-based metaheuristic optimization algorithm used in evolutional science