From: Not all biases are bad: equitable and inequitable biases in machine learning and radiology
Type of bias | Bias is rooted in | Example | Effects |
---|---|---|---|
Data bias | Low quality of datasets, due to | All types of biases can translate into machine biases and have effects on equity, accountability, and transparency | |
(1) quantitative misrepresentation of certain patient groups in datasets and/or | A set of training data for machine learning includes images of only (or mostly) males | ||
(2) qualitative misrepresentation of certain patient groups in datasets (e.g. wrong labeling of images) | A set of training data for machine learning includes images of socio-economically disadvantaged women as “healthy” controls, although some of them have diseases that were not diagnosed | ||
Cognitive bias | Features in the human processing of knowledge and cultural factors | Radiologists’ culturally informed stereotypes lead to a high rate of false negatives in breast images of socio-economically disadvantaged women |