DeepMind’s AI for Breast Cancer Screening
Screening mammography aims to identify breast cancer at earlier stages of disease, when treatment can be more successful. Despite the existence of screening programs wordwide, the interpretation of mammograms is affected by high rates of false positige and false negatives. DeepMind, a UK-based artificial intelligence company purchased by Google, has turned its sights to the problem. They have develped an AI model which can identify breast cancer from scans with fewer false positives or false negatives than experts.
DeepMind trained its AI using de-identified data from patients in both the US and the UK, and showed that it could reduce false positives by 5.7 percent and false negatives by 9.4 percent in the US. There was a smaller reduction of 1.2 percent and 2.7 percent respectively in the UK, which could be due to various reasons such as differences in imaging technology, disparity in data, or disparity in the underlying disease.
The AI only had access to the most recent mammogram of each patient whereas the human experts had the patient histories and prior mammograms to make their assesments. Despite this, the model was able to make screening decisions with greater accuracy than the experts and could be generalized to different populations.
In an independent study of six radiologists, the AI system outperformed all of the human readers: the AUC-ROC for the AI system was greater than that for the average radiologist by an absolute margin of 11.5%. They ran a simulation in which the AI system participated in the double-reading process that is used in the UK, and found that the AI system maintained non-inferior performance and reduced the workload of the second reader by 88%.
The developers of the AI emphasize that this is early stage research and that more studies and cooperation with healthcare providers will be required before the system is ready for widespread use. It would be interesting to see how the AI performs on rare forms of disease vs experienced practitioners as it probably does not have much training data to go off of. In any case, the potential for this system is great and could pave the way for clinical trials to improve the accuracy and efficiency of breast cancer screening.
