Theodore Cary, Abstract Co-Author: Nothing to Disclose

Emily Fox Conant MD, Abstract Co-Author: Nothing to Disclose

Peter H. Arger MD, Abstract Co-Author: Nothing to Disclose

Chandra M. Sehgal PhD, Presenter: Nothing to Disclose

To delineate breast mass boundaries in ultrasound images using the semi-automatic Plugged Leak Growing (PLG) algorithm.

The proposed algorithm grows a region of interest from a user-defined seed. Disk-shaped structuring elements of decreasing radii are iteratively fit around the seed to approximate the mass. As the disk size decreases, the ROI pixels leak into surrounding tissue from porous boundary sections. Leaks are identified by their areas relative to the disk area, and then plugged by masking them in the image. The process is repeated until all leaks have been plugged and the boundary has been approximated by 1-pixel radius disks. The algorithm was evaluated on digital phantoms and sonograms of 40 malignant and 40 benign biopsy-proven breast masses. The margins were identified by PLG in two trials using different user-defined seeds and compared with manual tracings using area correlation and overlap metrics.

For the 80 masses, the linear fit for the two semi-automated PLG trials has a slope of 0.94 and correlation coefficient (RČ) of 0.96, compared to 0.93 and 0.87 for manual trials. The correlation of both PLG trials with both manual trials has an average RČ of 0.86±0.07, suggesting that PLG is about as likely to agree with a manual boundary as a second manual boundary drawn by the same expert. In all PLG to manual correlations, the slopes of the fitted lines were less than 1 (0.92-0.96), indicating that PLG regions tend to be smaller than manually-traced ones.The results from the area overlap metrics are comparable to those found by area correlation. The ratio of area overlap to total lesion area, or true positive fraction, for benign and malignant regions found by PLG was 0.83±0.11 and 0.78±0.10, compared to 0.74±0.15 and 0.75±0.14 for manual tracing. The overlap ratio between manual and PLG boundaries was lower, 0.69±0.39.

Boundaries found by PLG are reproducible and detailed despite variations in user-defined seeds, correlate well and overlap reasonably with manual tracings, and could be of value in characterizing masses.Acknowledgements: Supported by NIH grants CA 85424 and CA 87526.

Cary, T, Conant, E, Arger, P, Sehgal, C, Semi-automated Boundary Detection of Breast Masses in Ultrasound Images.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4414074.html