Next ’19 Customer Innovation Series: American Cancer Society

Next ’19 Customer Innovation Series: American Cancer Society


MIA GAUDET: Hello, everyone. I’m Mia Gaudet from the
American Cancer Society. How about I tell
you about a story about how we’re
making cancer research advances in months rather
than years, compared to just a few years ago? Machine learning is
impacting cancer research across the continuum of cancer. Today I want to tell you
about a recent breast cancer project that leverages the
American Cancer Society’s long-term investment in
cancer data resources, together in partnership with
Slalom and Google Cloud. For over a century, the
American Cancer Society has dedicated its mission
to preventing cancer, saving lives, and diminishing suffering
from cancer through research, advocacy, education,
and service. As the strategic director
of breast cancer research at the American
Cancer Society, I lead a portfolio of research
in cancer prevention and prognosis. The more we learn
about breast cancer, the more we understand it’s
more than just one disease. So my colleagues and I
have dedicated our careers to better understanding the risk
factors for all types of breast cancer so that we can advance
prevention and treatment of different types
of breast cancer using personalized prevention
and treatment approaches. As cancer researchers,
we share a common problem with many of you. That is, we have a
wealth of data around us from many diverse sources. But it’s really understanding,
how can we yield new insights into this data? Data must be collected,
standardized, and analyzed so
that the results can be interpretable and useful. In our case, we
had over 20 years of lifestyle data and
cancer tissue specimens from women who were diagnosed
in the American Cancer Society’s cancer prevention
study II, which is one of the largest cancer
research projects in the US. Slides from these
tissue samples, such as the ones that
you can see held here, can provide new insights
about breast cancer. But the challenge is
how to identify and use clues beyond what a pathologist
can see under a microscope so that we can gain new insights
into the different types of breast cancer. Last fall, we initiated
a collaboration with machine learning experts
at Slalom and Google Cloud to unlock breast cancer insights
from the breast tissue samples themselves. So a first step– and one important, perhaps,
to your organization– is to understand each
other’s language and tools. I don’t speak
fluent computerese, and my colleagues at Slalom
don’t speak epidemiology. But it’s important
that we understand each other’s perspective
and insights. Working together,
we overcame barriers with the slides themselves. So there were many– there
was a lot of variability in the quality of
the slides, as well as the variability in
the color, and artifacts, such as the blue ink that you
can see at the bottom there. From there, we tiled
the images to improve computational efficiency, and
applied unsupervised machine learning at scale
using ML Engine. And this helped us to identify
patterns in the tissue images that you can see here, colored
by the different colors to represent the different
patterns that were detected in the tissue images. And so some of these
patterns that we detected were similar to what
the pathologist sees under the slides,
such as differences in the cell center caused by the
cancer, among other features. Finally, we grouped
these tiles, based on these different
patterns, into groups. And it would take a team
of pathologists years to identify these patterns. Even then, each of
those pathologists may not be able to
do this consistently. But a machine-learning
algorithm allowed us– a machine-learning algorithm
learned and identified these patterns without any
bias that humans bring, and also far exceeded the
pace of a pathology team. So any good scientist first
checks the quality control of their data. So the first step
was to examine what these different groups
might represent, from what we know
from the clinic. And in fact, one
of these pattern– one of these groups
showed us how– the fat cells in– showed us the fat cells
in the different cancers, as you can see from
the box in the blue. But other groups brought up
patterns that we didn’t expect, which is one of the benefits of
using an unsupervised approach. So in the past few months, we’ve
been matching these clusters with other data– surveys completed by these
women over the past 20 years, before and after they were
diagnosed with breast cancer, and we’re looking at these
clinical data, their lifestyles and other exposures
over their lifetime. And with these data, we
hope to better understand how to prevent different
types of breast cancer and identify the characteristics
of the most deadly of the breast cancer types. Our ultimate goal is to provide
individualized recommendations for cancer prevention, early
detection, and treatment. At the American Cancer
Society, we, like you, are passionate about
our success and what it means for the world. And like you, we feel the
urgency in our mission, the need to garner insights
from the data around us and then to act,
to give patients years rather than months. So finally, I’d like to
thank Slalom and Google for moving us closer
to this goal in ways that we couldn’t have imagined
even just a few years ago. Thank you.

Leave a Reply

Your email address will not be published. Required fields are marked *