Using a deep learning tool can help improve accuracy and reduce false positives during magnetic resonance imaging (MRI) scans to check for breast cancer, shows research led by New York University.
The deep learning algorithm was able to match the performance of a panel of radiologists with experience of diagnosing breast cancer and when the radiologist and algorithm performances were combined, overall accuracy improved.
Use of MRI scans to detect breast cancer is very accurate. This type of screening was previously limited to high-risk patients, but new research suggests medium and average-risk women can also benefit.
If a scan is interpreted to suggest a woman has breast cancer, she will be referred for a biopsy. Evidence suggests that for every biopsy that confirms cancer, two-to-four will not. False positives can be extremely stressful for women impacted by them and also waste hospital time and resources.
“As the number of patients undergoing breast MRI continues to increase, it is important to maintain high specificity and positive predictive value to minimize unnecessary biopsies and follow-up recommendations,” write the researchers in Science Translational Medicine.
Artificial intelligence algorithms, such as those using deep learning, can potentially help improve the accuracy of tests and diagnostic scans. In this study, Krzysztof Geras, an assistant professor in the Department of Radiology at New York University School of Medicine, and colleagues developed an algorithm to improve the accuracy of breast cancer screening.
Geras and colleagues first trained the model using 21,537 dynamic contrast-enhanced MRI (DCE-MRI) scans from 13,463 patients. The system was also validated using additional, independent datasets from both Poland and the U.S.
The algorithm was tested against a panel of five experienced radiologists and achieved similar diagnostic results. When the predictions from the radiologists and the algorithm were combined, the overall accuracy improved, suggesting that the tool could be a useful diagnostic aide to radiologists.
The researchers found that the algorithm could help reduce false positives and unnecessary biopsies in up to 20% of lower-risk patients (Breast Imaging Reporting and Data System 4) with potentially cancerous lesions.
“Clinically, our model may help personalize patient management, leading to a reduced number of unnecessary workup and biopsies, and may be better than a biopsy-all strategy in Breast Imaging Reporting and Data System 4 findings,” conclude the authors.