Public Release: 

Learning From Experience: New Pattern Recognition & Detection Helps Radiologists Analyze Digital Mammograms

Georgia Institute of Technology

An unusual new technique for pattern recognition and small object detection has successfully detected microcalcifications in digitized mammograms -- and holds promise for discerning other medical pathologies, manufacturing defects, and various objects in commercial, defense and Internet imagery.

Unlike most pattern recognition systems, this approach does not totally rely on the intervention of a human to extract and define a set of features for it. This new system is capable of efficiently searching a database of raw information to match patterns and detect objects.

"Instead of trying to extract a restricted set of features and train the classifier to look just for those features, we would have the expert user, in this case, the radiologist, directly build a database of known cancer indications encountered in the past," explained Dr. Christopher F. Barnes, a senior research engineer at the Georgia Tech Research Institute (GTRI). "We provide a quick-search software and firmware interface that allows this large database to be efficiently searched in near real time. Then, the radiologist's archive of past pathological cases essentially becomes a data classifier for processing new mammograms."

In its mammogram analysis, the software did not miss any microcalcifications in any of the digitized mammograms that were properly calibrated.

"In all of the data that's similar to our database data, this approach achieved nearly 100 percent detection with what appears to be acceptable levels of false alarms," Barnes said. "That's where you want to be with a mammogram -- we intend to provide a safety net for the expert-human analysis that may be susceptible to fatigue factors."

The system is not intended to replace radiologists or other medical or manufacturing professionals, Barnes says. It will merely suggest regions of mammograms or other data that should be given closer attention, based on past observations. Furthermore, the system provides confidence information related to the similarity between the imagery that is being analyzed and the past data archived in the database.

One of the advantages of GTRI's new approach is that as digital sensors/scanners achieve higher resolution, GTRI's classification system can exploit the increased resolution of the data. Eventually, some subtlety may be captured with the high resolution scanners that would be invisible at normal human-display resolution, or too voluminous to be displayed in a zoom-in format for human review.

GTRI's system would provide a high confidence pre-screen capability for this high resolution data and cue the radiologist to only the suspect regions of the mammogram. The high resolution data might permit a more precise discrimination between benign and malignant tumors imaged in mammograms.

"We don't know yet if our system can reduce unnecessary biopsies, but we do have reason to think our system can find subtle events that may be captured in future high resolution systems," Barnes added.

The system also might help radiologists reduce eye fatigue, increase productivity and possibly increase patient loads.

Partially developed with internal funding from GTRI, the system uses a set of structurally constrained templates generated with a proprietary design process that creates database addresses which are searchable in a computationally efficient manner.

Adding more examples of a particular pattern or object to the database usually does not significantly increase the amount of time require for a search, because the search is roughly de-coupled from the size of the database. The system also can mark cases similar to the example it is reviewing, so the user can make contextual evaluations. And the system can be configured to be trainable as needed by expanding the underlying database -- allowing the system to adapt to new cases.

The system is user friendly and keeps engineering costs low, Barnes says.

"All that is required is for a user to build the database, and then use some of our developmental tools to generate a software/firmware interface allowing efficient searches to be performed," he noted. "In fact, our goal is to have a system that is entirely user driven, all point- and-click."

In addition to other medical applications, the system can be applied to manufacturing quality control, preventive maintenance and defect detection -- identifying cracked fan blades in aircraft, for example. Among the types of data the system can search are acoustical signal wave forms, one-dimensional signals and acoustics, sonar, ultrasound and radar signatures.

"Law enforcement may be able to use it to find marijuana production fields in aerial photography, and oil companies could use it to do mineral analyses of spectral images," Barnes said. "It could have GIS applications for land use classification, and insurance companies could use it to identify potential landslide or fire-risk areas."

Barnes chose to test the system first on mammograms because breast cancer is a serious health threat. The American Cancer Society estimates that in 1997 some 180,200 new cases of breast cancer will be diagnosed among women in the United States, 1,400 cases among men. Also this year, 44,190 people, 99 percent of those female, will die from the disease -- it is the leading cause of cancer death among women 40 to 55.

The biggest challenge in testing the system on mammograms was collecting digitized mammography data -- radiologists weren't taking mammograms in digital form when Barnes began the project in 1991, and few people had taken time to digitize what they had. Dr. Debra Monticciolo, director of breast imaging at Emory University School of Medicine's Department of Radiology, provided initial data; the data Barnes later used was from National Expert and Training Centre for Breast Cancer Screening and the Department of Radiology at the University of Nijmegen, the Netherlands.

The next step is U.S. Food and Drug Administration approval, Barnes says, which means finding a research partner with enough interest to cooperate on extensive clinical studies. Suitable partners could include manufacturers of emerging digital mammogram scanners, medical research centers or insurance companies.


Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.