Deep learning expedites normal findings on ultrafast breast screenings

AI can safely and accurately identify healthy breast tissue on ultrafast breast MRI, negating the need for a radiologist’s closer look and, in the process, lowering cancer screening costs and widening patient access to breast MRI.

So conclude researchers at the University of Groningen Medical Center in the Netherlands whose report is running in European Radiology [1].

Xueping Jing, Peter M.A. van Ooijen and colleagues trained a deep learning model on 837 breast MRI scans from 438 women that were labeled at final interpretation as either normal (without suspicious lesions) or abnormal (with suspicious lesions).

Testing the model as tuned to a default probability threshold on an independent image set of 178 scans from 149 patients, the team found it brought back an AUC of 0.81. When they applied a high-sensitivity threshold, the researchers found their algorithm yielded a sensitivity of 98%, a negative predictive value of 98%, a workload reduction of 15.7% and a scan time reduction of 16.6%.

Further, these process improvements came at the modest clinical cost of 1 of 55 scans getting incorrectly read by the AI as negative.

“Although a conservative strategy was adopted, there were still false negative predictions,” the authors comment in their discussion. However, all the missed lesions were smaller than 10 mm, they point out, and “the relatively small size may be the main reason that the deep learning model did not detect them.”