Computer-Aided Diagnosis in Digital Mammography

Institution: Charles R. Drew University of Medicine & Science
Investigator(s): Jack Sklansky, Eng.Sc.D. -
Award Cycle: 1997 (Cycle III) Grant #: 3IB-0011 Award: $214,200
Award Type: IDEA
Research Priorities
Detection, Prognosis and Treatment>Imaging, Biomarkers, and Molecular Pathology: improving detection and diagnosis

Initial Award Abstract (1997)
We plan to create a computer system that will help radiologists to use large collections of digitized mammograms as aids in determining whether or not to recommend biopsies. Roughly two of every three breast biopsies recommended by a radiologist are unnecessary because they are benign (non-cancerous). The best available means of reducing the number of these unnecessary biopsies is a second opinion from a consulting radiologist. Often such a radiologist is not available or cannot be reached when needed. In these circumstances the proposed computer system will give the radiologist advice that is at least as reliable as that of a second radiologist, and possibly even more reliable. It will also give the radiologist a simple way of finding other mammograms that are medically and visually similar to the one being analyzed. This will enable the radiologist to diagnose mammograms more accurately and eliminate many biopsies of benign breasts. As a result, the proposed diagnostic system will lower the cost of second opinions, eliminate unnecessary surgical procedures, and reduce the degree of psychological stress among patients.

This research is made possible by "digitization" technology, which converts film mammograms into digital electronic data that can be processed by computer. Since about 1990 this technology has become so faithful to the original mammogram, so fast and so inexpensive to store, that several major hospitals have begun building large files of "databases" containing hundreds of digitized mammograms, along with the diagnoses and subsequent medical histories of the patients. Several of these databases have been made available to this project.

This research will be carried out at the Charles R. Drew University of Medicine & Science and the King/Drew Medical Center, the preeminent medical school and hospital in Southern California serving a primarily African-American and Latino community. A major benefit of this study will be the development of a computer-aided diagnostic system attuned to the particular patterns of disease, and the particular mammographic features, of this minority population. In addition, this study will result in improved service to an impoverished and medically underserved community. Further, the means of reducing the costs of unnecessary biopsies will take into account the ways medical costs are paid in this community.

Final Report (1999)
We built and tested a computer system to help radiologists use large collections ("databases") of digitized mammograms as aids in determining whether or not to recommend biopsies. We wished to measure the extent to which this "mapped database diagnostic system" helps the radiologist reduce the number of unnecessary biopsies without increasing the number of missed cancers. In this study we restricted the abnormalities to regions of interest (ROIs) exhibiting microcalcifications.

A "visual neural network" mapped each region of interest (ROI) in a database of proven mammographic ROIs as a dot on a computer screen. Each dot was labeled as either benign or malignant by a distinct color or shape. In addition the neural network constructed a "decision curve" splitting the screen into two regions - one for biopsy not recommended and the second for biopsy recommended. Annexing this display to an image retrieval system helped the radiologist to use case-base reasoning for diagnosis.

To make this diagnostic system fast and easy to use, a novel two-classifier system for automatically detecting and segmenting microcalcifications was devised, built, and evaluated. In a test on 86 ROIs, the automatic microcalcification detector-segmenter was compared to a manual microcalcification detector-segmenter. We found that the mapped database diagnostic system reduced the number of negative biopsies by 72% when using a manual microcalcification detector-segmenter, and by 67% when using the automatic microcalcification detector-segmenter. We concluded that the use of the two-classifier microcalcification detector-segmenter, coupled with a genetic selector of the features for the two classifiers, is a promising approach to the design of microcalcification detector-segmenters for use in the diagnosis of mammograms.

A test of the mapped database diagnostic system was carried out with a panel of four radiologists, using a database of ROIs in which microcalcifications are visible. In this test the sensitivity and the specificity of each computer-aided radiologist exceeded the sensitivity and specificity of each unaided radiologist and of the computer alone at p < 0.01. We concluded that radiologists interacting with this mapped database of proven mammographic ROIs are likely to achieve significant reductions in the number of benign biopsies and the number of misdiagnosed cancers when interpreting mammograms containing images of microcalcifications.

Using the mapped database system in mammographic diagnosis is likely to reduce the need for second opinions, reduce the number of unnecessary surgical procedures, and reduce psychological stress among the patients. Another benefit of this system is that it can be attuned to the patterns of breast disease in restricted populations.

Database and neural network help mammographers suppress benign biopsies and detect cancers. Hot Topics at 1997 RSNA Scientific Assembly
Index Medicus: Radiology
Authors: Sklansky J, Tao EY, Ornes CJ, Disher AC, Eisenman JI, Swartz JB
Yr: 1998 Vol: 206 Nbr: 2 Abs: Pg:205

Visualizing a database in mammographic screening. Scientific Program of 1998 Ann. Mtng. of Radiological Soc. of N. America, Chicago, IL, Dec. 1998.
Index Medicus: Radiology
Authors: Sklansky J, Tao EY, Ornes C, and Disher AC
Yr: 1998 Vol: 209 Nbr: P Abs: Pg:392