UCLA Radiology
Winter 2021
6
UCLA is currently exploring what role machine learning (ML)
and artificial intelligence (AI) can play in managing the high
daily workload for radiologists by aiding in the interpretation of
screening mammograms. “A reliable AI system could help with
workflow by efficiently triaging patients with suspicious findings on
screening exams and by helping mammographers reduce callback
rates,” explains Dr. Fischer. Decreasing callback rates reduces
patient radiation exposure, allays patient anxiet y and frees up
physician time to increase overall efficiency.
While computer-aided detection (CAD) systems have been
under development for decades, the first CAD software was not
approved for use by the FDA until 1998. CAD systems are very
different from the ML algorithms that are currently generating
a great deal of interest in many areas of radiology. While CAD
could highlight focal areas of increased breast tissue density and
microcalcifications, it has not proved to be impactful in helping
radiologists interpret image data or in increasing efficiency.
AI systems for mammography use deep convolutional neural
networks that learn how to classify image data. Such systems
are able to aid in breast cancer detection in a more nuanced
way than could earlier CAD systems by more adeptly handling
ambiguous data. Today’s AI systems evaluate mammography
images and assign numerical values to indicate the risk of breast
cancer. These AI systems provide a score for each finding on a
mammogram, calculating the probability of cancerous tissue for
each suspicious area of interest.
Dr. Fischer points out that current AI systems are not intended to
replace the human radiologist, but to serve as a smart assistant
in interpreting screening mammograms. The numerical results
of the AI system’s evaluation are available to radiologists in real
time as they review images, helping them better and more quickly
interpret the entire study.
UCLA is embarking on an extensive and multi-pronged research
program to evaluate the performance of AI in assisting the
interpretation of screening mammograms and contributing to
the clinical practice of screening mammography, as well as
how AI should be practically integrated into the high-volume
clinical workflow. The research is starting with a retrospective
study testing ML algorithms on 5,000 screening mammograms
performed at UCLA from 2010 to 2015. Various competing
ML algorithms have been tested by their developers, who
have reported their findings and have made claims about their
algorithm’s accuracy based on that data. “There is, however,
concern that the performance measures of these ML algorithms
using the vendors’ test cases may not be fully generalizable to the
screening mammograms performed at UCLA,” explains Hannah
Milch, MD, assistant professor of radiology, who serves as one of
the lead investigators in this research. “There might be differences
in patients’ diversity, breast density, medical and surgical history,
race, ethnicity and breast cancer risk.” After assessing how the
different ML algorithms perform on our own archived data set,
UCLA will install one of the systems and perform a prospective
clinical trial to fully evaluate how it performs in UCLA’s everyday
screening mammography workflow. “While there are some
interesting clinical trials coming out of Europe, we’re expecting
to be at the forefront of actually using and prospectively studying
AI in reading screening mammograms,” explains Dr. Milch.
Drs. Fischer and Milch and their colleagues are also thinking
about the practical adoption and future developments needed.
“Present AI systems do not look at prior films when they do their
interpretation, whereas the human radiologist does,” says Dr.
Fischer. “If the AI system could look at prior films and add that
information to what it detects in the present films, it will be more
useful, more accurate, and more able to diagnose very early
stages of cancer by detecting subtle digital imaging changes in the
breast tissue that may be difficult for the human eye to perceive.”
Other information that could be incorporated to improve future
AI systems includes demographics such as patient age, cancer
history, genetic information and even social determinants of health.
Foreseeing a day when AI will play an even larger role in triaging
screening mammograms, Dr. Fischer notes that many of the
breast imaging radiologist’s hours are currently spent assessing
healthy women. “With a robust, dependable AI system, we could
decrease the time spent on evaluating normals in the daily workload,
freeing us to spend more time on complex diagnostic exams, cross
sectional MRI exams, biopsies and other interventional procedures
— areas where AI cannot replace humans.”
“Screening mammography is the cornerstone of breast cancer detection,” says Cheryce Poon Fischer, MD,
professor of radiology, section chief and director of the Iris Cantor Breast Imaging Center. During the 12 months
ending in October 2021, over 38 million screening mammograms were performed in the U.S. “The sheer volume
of screening mammograms is staggering, requiring a large number of highly subspecialized radiologists for accurate
interpretation,” continues Dr. Fischer.
R
Hannah Milch, MD
Assistant Pro fessor of Radiolog y
Department of Radiological Sciences
David Geffen S chool of Medicine at UC LA
Cheryce Poon Fischer, MD
Professor o f Radiology
Section Chief of B reast Imaging
Director of Iri s Cantor Breast Imaging Cen ter
Department of Radiological Sciences
David Geffen S chool of Medicine at UC LA
Using Artificial Intelligence to
Interpret Screening Mammograms
