Surpassing old limitations in breast cancer screening, Google’s AI vies against radiologists to set new benchmarks.
Background
A team comprised of researchers from the United Kingdom and the United States recently published information highlighting that an artificial intelligence system designed by Google is better at detecting breast cancer via mammogram screenings than human radiologists. The new mechanism can enhance treatment going forward; it eliminates many common errors in cancer diagnosis.
“…radiologists miss about 20% of breast cancers in mammograms due to false negatives and false positives.”
The American Cancer Society reports that radiologists miss about 20% of breast cancers in mammograms due to false negatives and false positives. Mammography has been crucial in the screening and detection of breast cancer for decades. Still, even technological advancements over the years have not been able to remedy the high incidence of error.
Adding AI to the equation seems to be the solution.
Here, we must note that approximately 80% of cancers missed during mammography are visible but either go unnoticed or are incorrectly deemed to be benign. The aforementioned international research group came together to proffer a fix for this age-old situation.
Artificial Intelligence and Mammograms – The Study
To evaluate the AI system’s effectiveness in a clinical setting, they analyzed a large representative dataset from the UK and a large enriched dataset from the USA. It was found that the system was responsible for an absolute reduction of 5.7% and 1.2% in false positives and 9.4% and 2.7% in false negatives (in the USA and UK sets, respectively). The researchers independently studied six radiologists, all of which were outperformed. Thus, they concluded that Google’s system could provide better results. Astonishingly, the area under the receiver operating characteristic curve (AUC-ROC) for the mechanism was greater than the AUC-ROC for the average radiologist by an absolute margin of 11.5%. Also, the application of artificial intelligence to the double-reading process maintained non-inferior performance while reducing the workload of the second reader by 88%.
Alphabet’s DeepMind AI unit, which merged with Google Health in September 2019, was utilized in the study. Before its research work, Alphabet DeepMind gained a strong track record of creating high-performance AI systems adept at consistently outsmarting humans in video games. Its first medical application was focused on assisting doctors in the real-time analysis of anonymized eye scans, which sought early signs of diseases that might lead to future blindness in patients.
An Example of a banner with a button
Get the book for free
Most applications developed by Alphabet DeepMind rely on artificial neural networks, deep reinforcement learning, and machine learning as they analyze results over and over again to achieve optimal performance. The deep learning model used in the screening research was trained to identify breast cancers in tens of thousands of mammograms. In fact, the data was collected from 25,856 women in England between 2012 and 2015, and 3,097 in the United States between 2001 and 2018.
Coinciding and Confirming Findings
While addressing the accuracy of the study’s results, Connie Lehman, chief of the breast imaging department at Massachusetts General Hospital, confirmed that the findings were directly in line with ones from several other groups trying to use AI to improve cancer detection, including those published in her work. Lehman focused on the use of a deep learning model to triage a portion of mammograms as cancer-free, improving performance and workflow efficiency. Her utilization of the application revealed a potential reduction in radiologist workload and a significant improvement in specificity, without sacrificing any sensitivity whatsoever.
Lehman did concede that the computer-aided detection (CAD) programs did not make much difference in medical diagnosis. This is because most of the programs were only trained to identify what a human radiologist could see. AI computers, however, can spot cancers more effectively because they digest and learn from actual results from thousands upon thousands of mammograms. Accordingly, those systems have more capacity to identify cues that the human eye and brain may miss.
Conclusion
AI has grown incredibly. Now it takes on major roles like the diagnosis of diseases and the determination and planning of personalized treatment. The finding of this research work is an indication that AI-systems, if properly designed and maintained, are more proficient at catching cancer than humans. With the increased adoption of certain AI-based systems in diagnosis, more lives will be saved.
This post hits home for me because of breast cancer in my family. I dedicate this blog post to all the strong women who courageously and bravely endured their fight with breast cancer. I pray that advances in this field come swiftly so that we can reduce and eventually end its tirade.
