Most screening breast MRI exams are normal and do not need to be read by radiologists. An AI triage tool is able to correctly sort out almost 40% of these normal exams before a radiologist sees them.
Presenter: Kenneth G A Gilhuijs; University Medical Center Utrecht, The Netherlands
Source: RSNA 2021
Deep learning is feasible to identify normal breast MRI examinations with high certainty – it dismisses nearly 40% of normal scans without sorting out scans with malignant disease. This could reduce the workload of radiologists.
Extremely dense breasts are a risk factor for breast cancer. The DENSE trial has found that additional MRI-screening of asymptomatic women reduces the number of interval cancers in extremely dense breasts. First results have confirmed the detection of 79 additional breast cancers in almost 4,800 women, equaling 16.5 additional cancers per 1,000 screens. So far, all screening breast exams are read by radiologists. This also means that radiologists read a lot of screening MRIs that are perfectly normal. They do not require an expert radiologist’s review.
To reduce this kind of workload for breast MRI radiologists, Kenneth Gilhuijs, University Medical Center Utrecht, The Netherlands, and colleagues have evaluated an automated triage tool for screening breast MRI. Its aim was to dismiss examinations without lesions, while still identifying all examinations with cancer.
RSNA 2021 – Presentation: SSBR09-5
Kenneth GA Gilhuijs: AI-Triaging of Breast MRI For Radiological Review In The Screening Of Women With Extremely Dense Breasts
Creation and Validation
Their AI tool was set up in several steps:
- Firstly, MRI scans were separated into left and right breast, and orthogonal MIPs were created.
- A convolutional neural network was trained with the examinations from the first screening round. It learned how to separate breasts with lesions (BI-RADS 2-5) from normal breasts without lesions (BI-RADS 1).
- The algorithm then went through an eight-fold internal-external validation process with eight testing hospitals.
- The fraction of examinations that could be dismissed from radiological review was established at 100% sensitivity.
- Examinations from the second screening round were then used to validate the AI model.
Dismissed Lesions for Each BI-RADs Category
| BI-RADS | Lesions triaged (%) | Lesions dismissed (%) | p-Value |
| 2 | 85.0 (76.1 - 94.0) | 15.0 (6.0 - 23.9) | 0.001 |
| 3 | 88.3 (80.3 - 96.2) | 11.7 (1.8 - 19.7) | |
|
4 |
91.2 (86.8 - 95.6) | 8.8 (4.4 - 13.2) | |
|
5 |
100 | 0.0 |
As the number of BIRADS goes down, the number of lesions dismissed goes up, i.e., BIRADS 2 contains more lesions that can be dismissed.
Results
- For screening round one, the scans of 4,581 women were evaluated, as well as scans of 2,901 women for round two.
- The area under the ROC curve was 0.83 for the first screening round and 0.76 for the second screening round.
- 39.7% of scans without lesions were dismissed for round one, 41% for round two.
- Most importantly, none of the scans with a malignant lesion was dismissed for either round.
- All 79 examinations with cancers in the first screening round were detected, as well as all 20 examinations with cancer in the second screening round.
- Non-mass enhancing lesions were dismissed more often than mass enhancing lesions (26.4% vs. 11.3%).
- The performance in all eight hospitals was similar, both for screening round one and two.
- As the number of BIRADS went down, the number of lesions dismissed went up.
“Lesions classified as BIRADS-5 by radiologists were never dismissed by AI – this means that not a single cancer was missed,” underlined Gilhuijs. He added that the background parenchymal enhancement did not impact the ability of the method to triage between lesions that should be seen and that can be dismissed.
Conclusion
Deep learning is feasible to identify normal breast MRI examinations with high certainty, dismissing nearly 40% of normal scans without dismissing malignant disease. The AI tool reproduces well in the second screening round of the DENSE trial.
Presentation: AI-Triaging Of Breast MRI For Radiological Review In The Screening Of Women With Extremely Dense Breasts
Session code: SSBR09-5
Author: kf/ktg
Last update: 14 Feb, 2022
Presentation: AI-Triaging Of Breast MRI For Radiological Review In The Screening Of Women With Extremely Dense Breasts
Session code: SSBR09-5 Author: kf/ktg Last update: 14 Feb, 2022
