Bias in machine learning for computer-assisted surgery and medical image processing

Figures & data

Table 1. Some examples of methodology-centric elements of bias and their definitions.

Distribution biases

Unrepresentativeness: when the distribution of the evaluation data used in the paper differs from that observed in the clinic, due to the influence of human-centric elements of bias or other (possibly unknown) factors

Data selection bias: when the dataset itself captures only a portion of the relevant distribution due to experimental design, technology, or feasibility reasons

Annotation bias: when the annotations themselves include a consistent bias or source of error, such as simplification or differences between the in-plane and out-of-plane quality

Temporal/causal shift: when the distribution itself changes over time, possibly due to the influence of the system under investigation

Methodology selection biases

Model selection bias: when the model with the highest measured performance is selected from a group of measured models and said measurement performance is directly reported (i.e. no independent evaluation)

Metric/ranking selection bias: differences between the metrics used to evaluate a framework and the underlying ill-defined notion of clinical quality (an example of this is how one deals with class imbalance)

Data leakage

Feature leakage: when certain features are provided as input to an algorithm when they would be unavailable in practice temporal leakage: when information that is normally not available until after the machine learning task is provided as input for said task correlation leakage: when a feature is irrelevantly correlated with the prediction in the experiment but not in intended use

Train/test leakage: information from the evaluation (i.e. testing) dataset is used at least partially to develop the models which are then evaluated on said dataset normalization/imputation/augmentation leakage: when knowledge of the evaluation dataset is used to develop algorithms besides the primary one stacking/boosting/parameter leakage: when the evaluation set is used in the process of combining or selecting from multiple models or configurations