Page 17 - Layout 1

MEDICAL YEAR
IN REVIEW

a global stage launched a new era in ma- breast cancer cellularity, but also hormone gorithm would go to work, gathering precise
chine learning applications at SwRI. While receptor status of the cancerous cells. Breast data and information about that very spe-
we have developed machine learning algo- cancer is typically diagnosed using hema- cific patch of cells. That detailed informa-
rithms for other biomedical and health ap- toxylin and eosin (H&E) stain and assessed tion would mean faster patient diagnosis and
plications, such as markerless biomechanics visually for morphology, classification of the better treatment options to potentially save
in sports medicine and gait analysis to de- type and growth pattern of a tumor. Once more lives.
tect indications of cognitive decline, this morphology is assessed, tissues are stained
was our first time using machine learning with immunohistochemistry (IHC) to look AI Making a Difference
for cancer diagnostics. for predictive biomarkers. Visual diagnostics Artificial intelligence has become a part
naturally introduce human error. We are of modern life. Most of us interact with a
A Promising Future aiming to research the feasibility of using form of AI every day through streaming
Pathologists track tumor response to ther- our existing neural network expertise to clas- services, social media and in-home con-
apy by determining the percentage of tumor sify breast cancer into different groups nected devices. At SwRI, we have applied
cells in a particular area. Currently, this task based on hormone receptor status. machine learning to automotive, robotics
is performed manually and relies on experts Hormone receptor status is an important and defense technology. However, the
to interpret complex tissue structures. A de- prognostic and predictive tool for breast BreastPathQ Challenge presented a prob-
pendable automated method, like an algo- cancer patients, particularly in terms of lem that we had not previously tackled
rithm, produces faster results and more therapeutic response. As part of routine with artificial intelligence. Now that we
consistent data, while avoiding human error. pathology testing, breast cancer tissue is know what’s possible, we will continue to
Artificial intelligence and machine learning stained with IHC to observe certain bio- grow this capability. Our method contin-
approaches to medical imaging provide a markers, specifically estrogen receptor ues to garner enthusiasm and support
powerful tool to rapidly identify and quan- (ER), progesterone receptor (PR) and from the pathology community. As we ex-
tify cancer cells and guide treatment. human epidermal growth factor receptor-2 plore the potential for machine learning in
While our algorithm holds promise for all (HER2). The combination of these bio- cancer diagnostics, we expect more oppor-
types of cancer, for now, we are continuing markers informs diagnosis, therapeutic de- tunities to emerge to provide new health
to focus on breast cancer. As of 2016, breast cisions and risk of reoccurrence. Diagnoses care tools, and most importantly, improve
cancer is the most commonly diagnosed range from triple positive (positive to all patient outcomes.
form of cancer in females, with over three receptors) to triple negative. While
200,000 new cases annually since 2006, ac- these classifications are crucial to choosing Hakima Ibaroudene is Group Leader of R&D
cording to the Centers for Disease Control the proper treatment path, they are suscep- at the Southwest Research Institute.
and prevention. High diagnostic rates result tible to observer variability. Classifying
1
in high numbers of tests performed and breast cancer by hormone receptor status Citations
profiles collected. This testing surplus, com- with an algorithm would result in a more [1] “USCS Data Visualizations - CDC.”
bined with a predicted downturn in practic- definitive diagnosis and therefore, more tai- Centers for Disease Control and Prevention,
ing pathologists, could “negatively impact… lored treatment options for individual pa- Centers for Disease Control and Prevention,
health care providers’ abilities to deliver tients. Breast cancer is our starting point, https://gis.cdc.gov/Cancer/USCS/DataVi
more effective health care to their patient but patients with all forms of cancer could z.html.
populations.” These factors create a signif- potentially benefit from this capability. [2] Robboy, Stanley J, et al. “Pathologist
2
icant need for improved digital pathology Along with research to expand the algo- Workforce in the United States I. Develop-
tools to assist and automate parts of the tra- rithm’s capabilities, plans are underway to ment of a Predictive Model to Examine
ditional pathologist workload. make this new diagnostic tool a reality in Factors Influencing Supply.” Archives of
We are planning new research to expand pathology labs very soon. It could look Pathology, 5 June 2013, www.archivesof-
the algorithm’s capabilities to benefit both something like this – pathologists would pathology.org/doi/pdf/10.5858/arpa.2013-
pathologists and breast cancer patients. Our gaze at a monitor attached to a camera over 0200-OA.
goal is to use computer vision coupled with a microscope. The pathologist would select
deep neural networks to determine not only an image to study more closely and the al-

visit us at www.bcms.org 17