Quantitative bone metastases analysis based on image segmentation

Abstract

Preliminary evidence indicates that the fraction of bone containing metastatic lesions is a strong prognostic indicator of survival longevity for prostate and breast cancer. Our current approach to quantify metastatic bone lesions, called the Bone Scan Index, is based on an inspection of the bone scan, estimating visually the fraction of each bone involved and then summing across all bones to determine the percentage of total skeletal involvement. This approach, however, is time consuming, subjective and dependent on individual interpretation.

Methods: To overcome these problems, a semiautomated image segmentation program was developed for the quantitation of metastases from planar whole-body bone scans. The user is required to insert a seed point into each metastatic region on the image. The algorithm then connects pixels to the seed pixel in all directions until a contrast-dependent threshold is reached. The optimal threshold for cessation of the region growing is determined from phantom studies. On the images, lesion delineation and size measurements were performed by the algorithm. Each delineated lesion is associated with a bone site using pull-down menus. The program then computes the fraction of lesion involvement in each bone based on look-up-tables containing the relationship of bone mass with race, sex, height and age. These look-up-tables were obtained by multiple regression of the skeletal mass measurements in humans. The total fraction of skeletal involvement is then obtained from the individual fractional masses. For individual fractional mass, values given in International Commission on Radiation Protection Publication No. 23 were used.

Results: The bone metastases analysis system has been used on 11 scans from 6 patients. The correlation was high (r = 0.83) between conventional (manually drawn region-of-interest) and this analysis system. Bone metastases analysis results in consistently lower estimates of fractional involvement in bone compared with the conventional region-of-interest drawing or visual estimation method. This is due to the apparent broadening of objects at and below the limits of resolution of the gamma camera.

Conclusion: Image segmentation reduces the delineation and quantitation time of lesions by at least two compared with manual region-of-interest drawing. The objectivity of this technique allows the detection of small variations in follow-up patient scans for which the manual region-of-interest method may fail, due to performance variability of the user. This method preserves the diagnostic skills of the nuclear medicine physician to select which bony structures contain lesions, yet combines it with an objective delineation of the lesion.