Machine learning can help reduce a radiologist’s workload by identifying negative mammograms that do not need to be interpreted, according to new findings published in the Journal of the American College of Radiology.
“As examination volumes and time of interpretation increase with newer screening technologies such as digital breast tomosynthesis, radiologists will be under increasing pressure to deliver a timely service,” wrote lead author Trent Kyono, department of computer science at the University of California Los Angeles, and colleagues. “Because the large majority of mammograms a radiologist examines are negative, machine learning methods that triage a subset of examinations as negative with extremely high accuracy and refer the rest to a breast imager could significantly reduce the daily interpretive workload of radiologists, freeing up time to focus on more suspicious examinations and diagnostic workups.”
Kyono et al. developed a machine learning model called Autonomous Radiologist Assistant (AURA) using data from more than 7,000 women who underwent a diagnostic mammogram at one of six National Health Service Breast Screening Program centers throughout the UK. A convolutional neural network (CNN) was trained to read a single image from a patient’s mammogram and predict both the final diagnosis and all radiological assessments. The CNN could also estimate the patient’s breast density, patient age and other key details. A deep neural network was developed to fuse numerous views into a single view, predicting the patient’s diagnosis and making a recommendation on if a radiologist’s interpretation was needed.
“AURA is distinct from previous machine learning approaches in that it does not seek to replace human intervention, but rather assist radiologists by correctly classifying mammograms from patients with low risk of breast cancer, so that fewer mammograms need to be read by the radiologist,” the authors wrote.
For such a model to be effective, according to Kyono and colleagues, it needs to maintain a minimum negative predictive value (NPV) of at least 99%. AURA was able to maintain such a NPV and decrease the number of mammograms radiologists need to read by 34% in a theoretical setting where there is a 15% cancer prevalence. Most breast cancer screening facilities are expected to have a much lower cancer prevalence, however, and AURA showed the ability to decrease the number of mammograms radiologists need to read in a setting with low prevalence by more than 90%.
“Our findings suggest that the AURA system achieves this reduction by correctly classifying patients with attributes known to be associated with lower likelihoods of cancer, such as younger age and lower breast density, as breast cancer negative,” the authors wrote.
AURA-like software, if implemented, could potentially free up radiologists to have more time for focusing on difficult cases, the team added.
“The triage tool could be used in a number of ways, for example, prioritizing reading lists, allocating cases to readers of different experience, reading difficult cases at the beginning of day, and outsourcing lower-risk cases,” they wrote. “AURA provides a palatable solution for incorporating artificial intelligence into radiology that does not compete with the radiological practice but complements the workflow.”