FIGURE 2.

A summary of algorithm training and performance evaluation. (a) A subset of positron-emission tomography (pet) slices is used to train the algorithm. For each slice, 6 attributes (the “feature vector”) are calculated, and the reference threshold (“label”) is manually assigned by two radiation oncologists. Together, a feature vector and a label form a “labelled instance” (row in the table), and a set of such instances obtained from the selected subset of pet slices forms the “training set” (the table). During the training process, a learning algorithm uses the training set to learn how the label (the threshold) depends on the feature vector (the 6 attribute values). (b) Later, when given a new pet slice (from a “test set”), the 6 attributes are calculated and sent to the trained predictor, which returns the corresponding threshold. To evaluate the performance of the predictor, two radiation oncologists manually assign the reference threshold for this specific pet slice. Reference and predicted thresholds are then compared to evaluate the quality of the predictor. suv = standardized uptake value.