A model reads mammograms better than radiologists
On 1 January 2020 Nature published a Google Health and DeepMind model for breast cancer screening: on a UK and a US set it cut false positives by an absolute 1.2 and 5.7 points and false negatives by 2.7 and 9.4, and it outperformed each of six radiologists.
Why it matters
The model was tested in two countries with different screening rules, and one trained on UK data was shown to work on US data. In simulation it removed 88% of the second reader's work in UK double reading.
The UK set is representative, the US set enriched with cancers. The model's area under the ROC curve exceeded the radiologists' average by 11.5 points. The second-reader role was tested by simulation, not in the clinic, and the authors name clinical trials as the next step. In October 2020 Nature printed a correction and a comment on the paper with a reply by the authors; the record does not claim the model is used in screening.