Google Health pioneers breast cancer AI scanning
A research paper from Google Health published in Nature magazine has reported that machine learning, based on Google’s TensorFlow algorithm, can be used to reduce the level of false positives in breast cancer scans.
False positives occur when a mammogram scan is incorrectly identified as cancerous. A false negative occurs if it is wrongly diagnosed as not being cancerous.
The paper highlights the need for more clinical trials and the need for clinicians to adapt working practices as and when more AI (artificial intelligence) is used to support clinical decision making.
The mammogram scans represent images that can be analysed by an AI algorithm. Human-based image analysis of these scans is often challenging, according to the researchers, because cancer is often masked by dense breast tissue. This has led to computer scientists investigating the use of AI to analyse scans.
In the Google Health paper, based on training an AI algorithm to identify breast cancer using a large representative dataset from the UK and the US, the researchers reported an absolute reduction of 5.7% in false positives in the US dataset. The UK dataset exhibited a 1.2% reduction in false positive results.
The study also found that the system was able to reduce the false negative results in the US data by 9.4%. The UK dataset gave a 2.7% reduction in false negatives.
Although the mammograms were all based on scans using a system from a single manufacturer, the paper reported that the AI software outperformed human radiologists in accurately interpreting mammograms from screening programmes.
Using such AI systems to compliment human diagnostic work would require radiologists to adapt working practices. In particular, current systems that use computer aided diagnostics, tend to identify areas of a scan for radiologists to investigate further. If their examination of the scan reveals it is not cancerous, the scan is marked as negative for malignancy.
But, according to the researchers, if an AI-based system identifies an abnormal scan, that was not perceived as such by the radiologist, further investigation would be needed.
Researchers suggest that a clinical trial would enable clinicians to develop effective clinical practices to handle abnormal scans.
Jeffrey De Fauw, a researcher at Google’s Deep Minds, tweeted: “For AI health, we need well-designed clinical trials to validate performance, but this takes time.”
Describing caveats in how AI can indeed outperform human clinicians, De Fauw added: “Even given all these nuances, I’m still convinced that AI will have a strong positive impact on our lives, of which health will be an important aspect. I hope our work can contribute to that in the longer-term.”