FGI.

Top Line: You’ve likely heard by now about the Google/DeepMind AI system for mammography screening.

The Study: Here’s the publication describing that system’s performance on large mammogram test sets from the UK and US. The UK test included real-world screening mammograms from nearly 26,000 women where all reads were done by two radiologists (double-reading). The US test included just over 3000 women, and these were each read by a single radiologist (different ones for different reads). For interpretation, the AI system used the most recent previous mammogram for context. Compared to the first UK readers and the (single) US readers, the AI system significantly improved both the sensitivity and specificity for cancer diagnosis. In the UK dataset, the absolute improvements in sensitivity and specificity were +2.7% and +1.2%. In the US, those were +9.4% and +5.7%. What's more, the AI system was non-inferior even to the UK double-reads. Lastly, the AI system was put head-to-head with six radiologists on 500 randomly selected mammograms from the US dataset and outperformed them, too. Ok, so how could this be implemented? For one, it could immediately categorize patients in the low and high-risk ends of the spectrum to better triage human diagnostic efforts.

TBL: Google and DeepMind’s AI system for mammographic detection of breast cancer performs well—some may point out better than humans—in real-world cohorts of breast cancer screening patients. | McKinney, Nature 2020