AI may help detect breast cancer a year earlier than current methods
Key takeaways:
- An AI algorithm may help detect breast cancer among high-risk individuals 1 year earlier than usual.
- In the future, the algorithm may identify those at low risk to safely skip an annual screening.
An AI algorithm may help identify breast cancer on MRI scans 1 year earlier than current methods allow, according to research published in Academic Radiology.
Researchers trained the convolutional neural network AI model using MRI from 52,598 breasts.
They then used a retrospective dataset of 3,029 MRI scans from 910 patients (mean age, 52 years; range, 18-88) at high risk for cancer to fine-tune the model. The dataset contained 115 cancers diagnosed within 1 year of negative MRI.
The AI model detected cancers 1 year earlier with area under the receiver operating characteristic curve of 0.72 (0.67-0.76). Findings showed retrospective review by radiologists of the 10% of MRI the AI ranked as highest risk could have increased detection by up to 30%, researchers wrote.
A radiologist identified visual correlates to biopsy-proven cancers in 83 (72.1%) of cases. The AI model identified the anatomic region where the cancer ultimately would be detected in 66 (57.3%) of the 115 cases.
“It’s important to emphasize that radiologists are already extremely good at detecting cancers, and the fact that, in some cases, we can go back to the previous year and find cancer is actually good news for MRI technology,” Lukas Hirsch, PhD, a postdoctoral researcher in City College of New York’s Parra Lab, told Healio. “It means the technology we’re using is efficient at finding cancers. We’re just finding ways to leverage and use it more efficiently.”
Healio spoke with Hirsch about the motivation for developing the algorithm, the potential benefits it may offer and his team’s goals for effective implementation.
Healio: What motivated you to develop this AI algorithm?
Hirsch: One important motivation was to find a task that clinically has no solution yet. A lot of tools are being developed where the goal is to improve performance in a task we already have expertise in. That’s what we do in the lab — we develop neural networks that are meant to diagnose cancer in a current scan. That’s something radiologists are trained to do, so the goal there is to try to match or improve upon their performance. However, one thing radiologists do not do at the moment is predict whether cancer will develop in the next year. We don’t have a proper metric to evaluate that.
Healio: How did you conduct this study?
Hirsch: We had a retrospective database that included scans where radiologists found cancer and where there was a previous exam from that same patient that was deemed benign for whatever reason. We tested the AI tool on detecting malignancy prior to the positive finding. Of the cases that ultimately developed into cancer, we evaluated how many the AI tool was able to declare malignant beforehand.
Healio: What did you find?
Hirsch: It performed well. We had about 115 cases that did develop cancer. That sounds like a lot, but this is from a dataset of more than 50,000 MRIs, so it’s a very low number in general. Part of that is because radiologists are genuinely very good at identifying cancer, and there aren’t a lot of cases where cancer is missed. The algorithm did also predict a high probability of cancer in about a third of cases. This means for all the cases that did develop into cancer, there was an increase of 30% sensitivity — we might catch a third of these cancers a year early.
Healio: What are the next steps?
Hirsch: The way we developed this AI tool, it will flag suspicious cases to a radiologist, and the radiologist will then have to go over these cases again and reconsider the initial assessment. That also implies going over a lot of suspicious, benign cases that were correctly called benign. So, it is proof-of-concept that the AI tool is able to find cases that are generally suspicious, but we are still working on a clinical implementation that hopefully will require less work for radiologists, rather than more.
Healio: Is there anything else you’d like to mention?
Hirsch: We also plan to evaluate patients the AI predicts to be very low risk for future years. That would mean we could adjust the screening intervals based on personal risk. Currently, we’re focusing on patients at very high risk, where the AI thinks there might already be a malignancy and it needs re-evaluation. However, there’s the other end of the spectrum, where the AI says there is so little risk that patients might be able to skip their next screening interval. That’s something that radiologists also don’t necessarily do at the moment — it’s a 1-year fixed screening exam. So, we are hoping we can also use AI to identify patients at very low risk who can come every 2 or 3 years without risk for cancer being missed.
Reference:
For more information:
Lukas Hirsch, PhD, can be reached at lukashirsch@gmail.com.
Collapse
Healio Interviews
Read more about