Google has announced it has developed a machine learning system that can detect breast cancer from mammography scans faster and with greater accuracy than radiologists. In a recent study, the system outperformed radiologists and detected cancerous nodes that had been missed by skilled radiologists with a much smaller number of false positives and false negatives.
Breast cancer is most common form of cancer in the United Kingdom. More than 55,000 women are diagnosed with breast cancer in the United Kingdom each year. Around 1 in 8 women will develop breast cancer at some stage of their lives. Breast cancer is treatable, but the effectiveness of treatment is largely dependent on the stage of the cancer. Once the cancer has metastasized, treatment is far less effective. Early detection is therefore critical but this can be a major challenge, even with breast cancer screening programs.
Screening for breast cancer often results in false positives, where cancer is mistakenly diagnosed. This naturally causes considerable stress for patients. Worse are false negatives, where cancerous nodes in mammography images are missed. That can delay diagnosis and treatment, potentially past the point where treatment is likely to be effective. Screening programs also place a great strain on radiologists, of which there is a short supply. That can further delay detection.
Machine learning systems could potentially be used to decrease the workload of radiologists and improve the accuracy of diagnosis, especially at the early stages of the disease when it is most treatable but also most difficult to detect.
The machine learning system was developed by Google Health in 2017 to identify metastatic breast cancer from lymph node specimens. Google Health partnered with the DeepMind, Cancer Research UK, the Royal Surrey County Hospital and Northwestern University and used a dataset of deidentified mammograms from more than 91,000 women from the United Kingdom and United States to train its system.
Google Health recently put its machine learning algorithm to the test in a clinical setting using a separate data set of 28,000 de-identified mammograms from U.S and UK hospitals and compared the detection rate, false positives, and false negatives with those from skilled radiologists. They found their system was faster and more accurate than skilled radiologists at diagnosing breast cancer.
In that test, the machine learning system reduced the false positive rate by 5.7% for U.S women and 1.2% for British women. There was a 9.4% reduction in false negatives in U.S women and a 2.7% reduction in false negatives in British women. In another test the system was pitted against six radiology experts and it outperformed all six. The latter evaluation showed that in a clinical setting, where there is a double reading process in which the AI system participated, there was a reduction in the workload of the second radiologist of 88%.
Google Health also demonstrated that their model could generalize to other health systems. They trained their system only using a representative data set from UK women and evaluated the system on a data set exclusively of US women. In that test, the system resulted in a 3.5% reduction of false positives and an 8.1% reduction of false negatives.
In these tests, Google’s machine learning system also received less information than was available to radiologists, who had access to prior mammography images from patients and their medical histories. Google Health’s machine learning system only had access to a patients’ latest mammography images, yet it still outperformed human breast cancer specialists.
The positive findings from these evaluations will pave the way for clinical trials which should help to further improve the accuracy and efficiency of the system. It is hoped that the machine learning system can be used to reduce wait times, improve the accuracy of diagnoses, speed up the provision of treatment, reduce stress for patients, and improve patient outcomes.
You can read more about the study in the paper – International evaluation of an AI system for breast cancer screening – was recently published in the journal Nature. DOI: 10.1038/s41586-019-1799-6