<title>Preliminary evaluation of an "intelligent" mammography workstation</title>

Abstract

We are developing computer-aided diagnosis (CAD) schemes for the detection ofclustered microcalcifications and masses in digital mammograms (1-4). Here, CAD refers to a diagnosis made by a radiologist who uses the computerized analyses of radiographic images as a "second opinion" The radiologist would make the final diagnostic decision The aim of CAD is to improve diagnostic accuracy by reducing the number of missed diagnoses. In this preliminay evaluation, 30 clinical cases from December 1991 having a focal mammographic finding were analyzed. THE "INTELLIGENT" WORKSTATIONWe have implemented our computer-aideddiagnostic schemes on a high-speed computer for the quantitative analysis of radiographic images. This is essential for use of on-line CAD in clinical practice. The components of the "intelligent" workstation for CAD include a high speed computer (IBM RISC 6000 Series 560 Powerstation) and a laser scanner (Konica film digitizer LD 4500). The computer includes 512 MB main memory and four. 1 GB disks, and has a CPU clock rate of 50 MHz. Conventional mammograms are digitized in approximately 30 seconds (per film) to a matrix size of 2048 by 2580 with 10-bit quant.ization. For a given mammographic case, four films are currently digitized --the left and right craniocaudal (CC) views, and the left and right medio-lateral-oblique (MLO) views. Results of the computerized analyses are output to the CRT screen of the workstation with offset arrows indicating the position of the suspected abnormality.Our computerized scheme to locate clustered microcalcifications on mammograms operates on each mammogram separately. With our method, a digital mammogram is processed by filtering techniques in order to improve the signal-tonoise ratio of the microcalcifications in the image (5-8). Next, both a global gray-level thresholding procedure and a locally adaptive gray-level thresholding procedure are used to identify potential signals (microcalcifications) in the noise background. The segmented image is then subjected to feature-extraction techniques to remove signal that likely arise from structures other than microcalcifications. These techniques include an area filter, an analysis of the spatial-frequency components and a clustering filter (8-12). In addition, an artificial neural network can be used to eliminate more false-positive detections (13). Currently, on the workstation, the CPU time for the detection of clustered microcalcifications is approximately 23 seconds per image.Our computerized scheme for the detection of mass lesions operates on pairs of mammograms, e.g., the left and right CC views or the left and right MLO views. The method is based on the deviation from the normal architectural symmetry of the right and left breasts and uses a bilateral-subtraction technique to enhance the conspicuity ofpossible masses (14-18). After the right and left breast images in each pair are aligned, a nonlinear bilateral-subtraction technique is employed that involves linking multiple subtracted images to locate initial candida...