In breast cancer screening programs, such as the one running in the Netherlands, a high volume of mammography data is acquired. Radiologists have to evaluate hundreds of images every day with precision and often under time pressure. Artificial intelligence (AI) offers radiologists smart support: “It’s like having an additional colleague at the press of a button,” says radiologist Ritse Mann of Radboud University Medical Center, Nijmegen, the Netherlands.
The radiologists at the breast clinic at Radboud University Medical Center evaluate a steady stream of images every week with the help of Transpara (ScreenPoint Medical). This is an application for interactive decision support that is perfectly connected to the syngo.Breast Care reading solution from Siemens Healthineers. For these radiologists, their AI tool offers more than a second check – as an additional radiologist at the hospital would do. With the software’s score feature, mammograms are evaluated for the chance of malignancies even before a radiologist looks at them. Scores of 1 to 5 indicate a very low risk, while a 10 signals the highest chance of a malignant anomaly. Mann: “Let’s say the score is 3, then I know pretty much for certain that the chance of finding a visible cancer is virtually zero. I do take a look at the images, but it doesn’t yield much information.”
Especially by using this score level, Mann sees great potential for time savings because the AI tool could be the first “radiologist” to look at the images. This means that in cases of a very low score, a second reader – as is required in many breast cancer screening programs – could become superfluous. Therefore, it is a tool that appears to have great potential, above all, for screening. “Because screening is becoming more precise, fewer women have to come back to the hospital for further analysis while the same sensitivity is maintained.” This is why AI can help to provide better patient care. According to Mann, however, the impacts are particularly tangible on a broader level, mainly for the imaging institute, and not so much at the individual level.
AI-driven case scoring holds great potential
Another benefit AI brings with this software is the built-in decision support. This allows radiologists to make more precise evaluations of lesions and calcifications. The tool works interactively. If a radiologist sees an anomaly on a mammogram or tomosynthesis, he or she can click on the suspicious region. The software then shows, based on a score of 1 to 95, how high the chance is that a malignancy is present. Mann: “If a finding scores very low, I know that it is likely to be a benign anomaly and I can ignore it.”
The performance matches that of a good radiologist.
AI speeds up tomosynthesis evaluation
For diagnostic imaging, the breast clinic at Radboud University Medical Center uses tomosynthesis to acquire 3D images of the breast. This technique provides a higher depth of resolution, and therefore tissue separation, than conventional 2D mammograms, but it also generates a greater number of images for evaluation. Mann: “As opposed to one mammogram, you have sixty tomosynthesis slices. In practice, that means that we need twice as much reading time to evaluate a ‘tomo’.”
This “extra time” is also partly made up for by using the decision support tool in the AI software. It screens the images like a virtual radiologist, which enables human radiologists to make their evaluations faster – 15 to 20 percent faster, as various studies have shown. As a result, radiologists have more time for more complex cases.
Artificial intelligence from the patient’s perspective
Artificial intelligence is a relatively new field in breast care, with few companies active in it so far. Mann: “Transpara is one of the few genuine AI applications on the market that can work with both mammography and tomosynthesis. Many other technologies are still in the research phase and are not yet available in practice.” This is despite the fact that AI can play a very big role in breast cancer care. “Especially in screening, AI can reduce costs. For example, if it indicates that a woman has a low risk of breast cancer, then you have to wonder whether you still need an actual radiologist for the second check. The situation now is that two radiologists look at every mammogram no matter what. So, there is a lot of potential to save on staff.” In addition, there is also potential in AI as a teaching tool for junior radiologists, who could read and evaluate cases and then check their result against the tool’s decision support feature.
Nevertheless, Mann realizes that many steps must be taken before that. “From an ethical standpoint alone, it is still hard to say to a woman: the computer has evaluated the images and confirmed that there is nothing to worry about. The patient expects a ‘real’ doctor to look at the image. At least for now. But AI learns extremely fast. The more images we can use to let the system fine-tune itself, the higher the quality of the system and the care will be. Incidentally, it does raise new questions from the standpoint of privacy. If we can find answers to them, the quality and efficiency of breast diagnostics will make even greater leaps forward.”