AI Rivals Radiologists in Judging Breast Density

Dense breast tissue appears opaque in mammograms and can hide tumors. Scientists at NYU used 200,000 anonymized mammograms to teach a deep learning algorithm to classify breast density.

Computer scientists at New York University Center for Data Science together with radiology researchers from NYU School of Medicine trained an artificial intelligence algorithm to classify mammography images according to breast density.

They found that the AI categorizes mammograms as well as human readers do.

Breast density is defined in terms of the proportion and distribution of fibroglandular (dense) tissue in the breast. There are four density categories.

Density matters because it may be a risk factor for breast cancer. Laws in 31 states require that breast cancer screening patients determined to have dense breasts be notified that they may benefit from further screening.

“Dense breast tissue reduces the effectiveness of mammography because it has a ‘masking effect’ and will hide an underlying tumor,” said Linda Moy, MD, co-author of the study, professor of radiology at NYU Langone Health who specializes in breast imaging, and member of CAI2R. Dr. Moy added that “Dense breast tissue is common and normal.”

To create an automatic classifier of breast density, scientists used a method known as deep learning. In deep learning, an input (e.g. an image) is broken down and passed through layers of functions to produce an output (e.g. recognition of a face in the image). The output is then evaluated, and the network adjusts its internal functions in order to achieve higher accuracy for the next input, before the process repeats. Such networks, used widely in image and speech recognition, need enormous sets of data to train on.

The NYU study included more than 200,000 real anonymized mammograms—a dataset "approximately two orders of magnitude larger" than those used in previously published studies of medical image analysis, according to the authors. Researchers trained the network on 80 percent of the mammography cases, and validated the results with another 10 percent. After training and validating the network, scientists tested it on the remaining 10 percent of cases.

Authors measured how well the model classified mammograms by comparing its outputs to the original clinical classifications done by human radiologists. The scientists also compared assignments made by the model to new analyses performed by three human readers: a medical student, a radiology resident, and an attending radiologist.

The results were close.

"We find that our model can perform this task comparably to a human expert," the authors write.

The scientists also point out that their model offers "quantitative, reproducible prediction, while there is often poor intra-reader and inter-reader correlation in qualitative assessment of breast-density tissue."

“Our work is a good example of how much machine learning can support the work of radiologists,” said Krzysztof Geras, PhD, postdoctoral researcher at the Center for Data Science and co-author of the study. Dr. Geras added that other elements of breast cancer screening evaluation are “still a challenge for existing machine learning algorithms.”

But this element may be ready for translation.

“There is an unmet need to have a more precise, automated assessment of breast density,” explained Dr. Moy, adding that AI can also discern the pattern of dense tissue, further informing the overall evaluation of cancer risk. A density classifier, if made fast, affordable, and comprehensive “would be readily incorporated into clinical workflow,” said Dr. Moy.

Related study:
Wu N, Geras KJ, Shen Y, Su J, Kim SJ, Kim E, Wolfson S, Moy L, Cho K. Breast density classification with deep convolutional neural networks.
arXiv:1711.03674 [cs.CV]


Latest Updates

04/05/2021 - 09:00
03/24/2021 - 17:06

Philanthropic Support

We gratefully acknowledge generous support for radiology research at NYU Langone Health from:
• The Big George Foundation
• Bernard and Irene Schwartz

Go to top