FY2010 - Oak Ridge National Laboratory

Director’s R&D Fund—
Ultrascale Computing and Data Science
for Medicine using SEmi-supervised Learning (DAMSEL),” we (1) developed an analytical, automated
learning framework and tools for processing multimodality medical data (text and images) for the purpose
of data mining and assessment; (2) improved performance, portability, and scalability of this
computational framework by leveraging available intelligent software and hardware computing resources
and adding functionality to the system; and (3) conducted preliminary validation of the performance on
medical data. These aims were accomplished in the context of two biomedical applications: breast cancer
(mammography—imaging, pathologies, text reports) and Abdominal Aortic Aneurysm (AAA) (using 3D
imaging modalities, such as MRI, in addition to surgical and clinical notes from the patient record). In
addition, initial exploration of data for traumatic brain injury (TBI) [using Joint Theater Trauma
Registries (JTTR) and related TBI source data] was conducted.
Mission Relevance
DAMSEL has the ability to facilitate the development of more powerful analytical tools by leveraging all
of the data in a more effective manner than present approaches permit. These technologies are at the
forefront of systems medicine application development and require computationally intensive
environments for processing. Potential sponsors for follow-on funding include the National Institutes of
Health (NIH) and the Department of Defense (DoD), which have open program announcements and
planned initiatives in these areas. For example, NIH’s funding opportunity PAR-09-218, Innovations in
Biomedical Computational Science and Technology, will support research in tools for data acquisition,
archiving, querying, retrieval, visualization, integration, and management; platform-independent
translational tools for data exchange and for promoting interoperability; and analytical and statistical tools
for interpretation of large data sets. This project is also consistent with such DOE programs as
Mathematical, Information, and Computational Sciences, KJ.01.00.00.0; Computer Science,
KJ.01.01.02.0; and Computational Partnerships, KJ.01.01.03.0. Other DOE programs that will benefit
include national security, intelligence, and biosurveillance applications.
Results and Accomplishments
The project created a multimodal learning framework and tools for the analysis of mammography images
and reports and also for abdominal aortic aneurysm (AAA) images and reports. We developed a semisupervised
machine learning framework that integrates the text and image modalities by transforming an
image feature vector produced through image processing to a lower dimensional space that is smooth with
respect to the problem-specific similarities described in the text reports. The DAMSEL project provides
support for combining image and text modalities in previously unavailable ways, but the general
framework is also generic enough to support the combination of any number of different modalities that
represent different views of the medical problem. The effectiveness of the framework when the secondary
modality set is engineered to consist of features representative of the target problem can be dramatic and
has been demonstrated via improvement over state-of-the-art results.
The text analysis work for breast imaging produced the following key new capabilities: (1) a geneticalgorithm-based
approach to identifying reports of abnormalities; (2) a genetic-algorithm-based approach
to identifying key phrase patterns in the language used for the mammography domain; (3) a classifier for
mammography documents; and (4) a temporal analysis approach for examining and finding key phrase
patterns that behave as precursors to a future event in mammography patients. The key-phrase patterns
represent a highly effective set of features for creating cancer-related dichotomies in the data and support
the discovery of a valuable image-processing manifold through the framework. Additional work in year 2
included new analyses for the AAA data. This effort resulted in (1) a searchable index of the
mammography reports that includes specific information such as labels, date, key-phrase patterns
(s-grams), and anonymized patient IDs and can be presented in human- or machine-readable (xml)
format; (2) a patient-centered approach that uses all reports per patient and their timestamps to facilitate
exploratory temporal data analysis and visualization using information retrieval and classical statistical
67