Type:
Educational Exhibit
Keywords:
Artificial Intelligence, CT, MR, Computer Applications-General, Segmentation, Technology assessment, Education and training
Authors:
P. Gray, S. O'Hanlon, A. Gupta; WA/AU
DOI:
10.26044/ranzcr2021/R-0345
Imaging findings OR Procedure details
Terminology
The authors found a lack of standardisation within studies when describing their methodologies and what AI they have used, if it is described at all. Some studies report software used, some the coding environment and some either have developed their own code, which they may or may not make available to readers. If we think of the methodology as describing how a study has been performed so it could be replicated for the same results, then most AI study methodologies are currently lacking sufficient detail.
While something of an oversimplification, it can be helpful to think of such methodologies in a similar way to how we learnt about DNA. The 'code' is the instructions, this can be thought of like DNA. The coding structure is a particular way the code has been written or organised, like a double helix. The coding environment is a way to make the computer read the code and act on it, like the ribsome reading the DNA (or mRNA) to make an amino acid. The software is all of the above bundled together.
Definitions
- Decision Tree: A learning technique which uses a tree-like structure composed of several nodes (a root node, internal nodes, and leaf nodes) to predict the target variable from the features (independent variables)
- e.g. feathers? y= bird n= dog
- Random Forest: An ensemble learning method in which random subsets sampled with replacement from the training data are used to build multiple decision trees
- Convolutional Neural Network: A deep learning algorithm, extracts features from an image in 'layers'. The layers of a CNN consist of an input layer, an output layer and hidden layers that may include multiple convolutional layers, pooling layers, fully connected layers and normalisation layers.
- Multi-Atlas Approach: A segmentation technique wherein an 'atlas' or gold standard references are used, mapped onto the new image and then the labels are transferred.
- Thresholding: Where known boundaries are defined to split data into groups.
- e.g. using hounsfield units to classify different structures in a CT image
- Clustering: grouping together based on similar features.