Get started

Please fill in the form below and an Ezra representative will contact you within 1 business day.

Notify me

Please fill in the form below and we will be in touch when we launch in your city.

Blog / Artificial Intelligence, Breast Health, Cancer, Early Detection, Healthcare Trends

For Radiologists, AI Can Help Lighten Mammography Burden

Jun. 12 2019 by Sheherzad Raza Preisler Blog Editor, PR & Social Media Coordinator
For Radiologists, AI Can Help Lighten Mammography Burden

Researchers at UCLA and the University of Cambridge collaborated to develop a deep learning-based model known as Autonomous Radiologist Assistant (AURA); their goal was to create a program that can weed out “normal” mammograms, enabling radiologists to shift their focus towards analyzing abnormal and ambiguous ones.

A proof-of-concept study published in May showed that AURA has the potential to lower the amount of negative mammograms read by radiologists while preserving a 99% negative predictive value–or 99% of those with negative mammograms truly did not have breast cancer.

To train AURA, the research team used a data set comprised of mammograms from more than 7,000 women in six National Health Service Breast Screening Program centers in the United Kingdom. And unlike other, similar studies, this one did not seek to potentially replace the role of radiologists using convolutional neural networks (CNNs). Instead, the team attempted to create a “triage system” in which AURA autonomously diagnosed mammograms as negative, while leaving the rest of the studies to be interpreted by a radiologist.

CNNs work by mirroring the way radiologists read mammograms: first, they extract imaging features, then predict a diagnosis and radiological assessments for all four mammographic views. Afterwards, a deep neural network comes to a patient-level diagnosis once it has examined the CNN’s predictions for the mammography views as well as other, non-imaging features. Finally, the CNN will recommend whether it thinks a radiologist should read the mammogram.

The study looked into AURA’s performance in two situations: diagnostic and screening environments. In the diagnostic setting, there was a 15% prevalence of cancer, while in the screening setting, there was a 1% incidence of cancer. AURA was able to reduce the number of negative mammograms in need of radiological examination by 34% and 91%, respectively.

To the researchers involved, AURA can be helpful in a number of ways. It can potentially:

  • Help screen mammograms
  • Decrease the number of mammograms radiologists have to read
  • Reduce the likelihood of misdiagnoses because radiologists will have more time to spend on complex cases

While the study and possibilities seem promising, additional studies will have to be done to further evaluate AURA’s safety and efficacy.