The software being tested comes from Vara, a German-based startup that also led the research. The company’s AI is already used in more than a quarter of German breast cancer screening centers and was introduced earlier this year to one hospital in Mexico and another in Greece.
The Vara team, with the help of radiologists at Essen University Hospital in Germany and Memorial Sloan Kettering Cancer Center in New York, tested two approaches. In the first, the AI works alone to analyze mammogram images. On the other hand, the AI automatically distinguishes between scans it considers normal and those that cause concern. It deals with the latter for a radiologist, who will review them before looking at the AI’s assessment. The AI will then issue an alert if it detects cancer when the doctor doesn’t.
“In the proposed AI-based procedure, almost three-quarters of the screening studies do not need to be reviewed by radiologists, and improve overall accuracy.”
Charles Langlotz
To train the neural network, Vara fed AI data from more than 367,000 radiographs — including the radiologist’s notes, the initial assessment, and information about whether the patient ultimately suffered a stroke. cancer – to learn how to place these scans into one of three groups: “normally confident”, “Not confident” (where no predictions were made), and “confident cancer” .” The conclusions from both methods were then compared with the decisions radiologists actually made initially on 82,851 mammograms obtained from screening centers that did not contribute images. Scanned images are used to train the AI.
The second approach – doctors and AI working together – detected breast cancer 3.6% better than a single doctor working alone and gave fewer false alarms. It accomplished this while automatically skipping scans it classified as perfectly normal, representing 63% of all mammograms. This powerful streamlining can cut the workload of radiologists.
After screening for breast cancer, patients with a regular scan will be sent, while an abnormal or unclear scan will trigger further testing. But radiologists are checking mammograms miss 1 out of 8 cancers. Fatigue, overwork, and even the time of day all affect how well radiologists can identify tumors when they look at thousands of scans. Visually subtle signs are less likely to cause alarm, and dense breast tissue — mainly found in younger patients — makes cancer signs harder to notice.
Real-world AI radiologists are required by German law to review every mammogram, at least glancing at what the AI calls good. The AI still lends them a hand by pre-populating reports of normal-labelled scans, although the radiologist can always reject the AI’s call.
Thilo Töllner, a radiologist who heads a German breast cancer screening center, has been using the program for two years. He sometimes disagrees when the AI classifies scans as normal confidently and manually fills in the reports to reflect a different conclusion, but he says that “the norms are almost always normal. ” Mostly, “you just press enter.”
A mammogram that the AI has labeled as unclear or “confident cancer” is referred to the radiologist – but only after the doctor has made an initial, independent assessment.
