Itamar Arel

UTK electrical engineering and computer science

The human brain is capable of some truly formidable tasks. One of those is the ability to assess images, text, and data all at the same time and, with substantial training and experience, to arrive at a very high level of comprehension in order to make critical judgments.

But humans also need to sleep, and so even if their working hours are focused and effective, they still need help to handle the amount of information required to provide effective health care to a large population.

Both cost and scale make it essential that health information technology assist in accurate, judicious, cost-effective diagnoses. But, mimicking the human capacity for studying high-dimensional inputs—such as large images and videos—and identifying particular medical content and patterns within such inputs present a challenge for the field of machine learning.

The JDRD research of Itamar Arel promises to deliver a breakthrough in the field of medical image analysis using a family of machine learning systems known as Deep Belief Networks (DBN). These are derived from recent neuroscience research that has discovered how the cortex, which is the seat of most cognitive functions of the brain including pattern recognition, captures spatial and temporal information in observed data by maintaining a hierarchical neural network. Through modeling these processes with DBN, computer-aided diagnosis can assist highly skilled medical practitioners in more efficiently determining patterns of interest and significance.

JDRD project: A semi-supervised learning system using deep belief networks for multi-modal medical data analytics
LDRD project: Data analytics for medicine using semi-supervised learning (DAMSEL), Barbara Beckerman.

The corresponding LDRD research of Barbara Beckerman and her team in the ORNL Biomedical Science and Engineering Center (BSEC) is an alternative machine learning scheme known as DAMSEL, or Data Analytics for Medicine Using Semi-Supervised Learning. That research aims to provide a computational framework for both assessing and “mining” data in medical information systems by examining statistical properties and regularities within it.

The two learning schemes—DAMSEL and DBN—will allow for testing each against the other in order to both validate and support their respective approaches. Their test bed is mammography (breast cancer analytics) for which imaging, pathology data, and textual reports exist.