QUANTITATIVE BONE METASTASES ANALYSIS BASED ON IMAGE SEGMENTATION

Preliminary evidence indicates that the fraction of bone containing me tastatic lesions is a strong prognostic indicator of survival longevit y for prostate and breast cancer. Our current approach to quantify met astatic bone lesions, called the Bone Scan Index, is based on an inspe ction 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 consum ing, subjective and dependent on individual interpretation. Methods: T o overcome these problems, a semiautomated image segmentation program was developed for the quantitation of metastases from planar whole-bod y bone scans. The user is required to insert a seed point into each me tastatic region on the image. The algorithm then connects pixels to th e 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 an d size measurements were performed by the algorithm. Each delineated l esion is associated with a bone site using pull-down menus. The progra m 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 re gression of the skeletal mass measurements in humans. The total fracti on of skeletal involvement is then obtained from the individual fracti onal masses, For individual fractional mass, values given in Internati onal Commission on Radiation Protection Publication No. 23 were used. Results: The bane metastases analysis system has been used on 11 scans from 6 patients. The correlation was high (r = 0.83) between conventi onal (manually drawn region-of-interest) and this analysis system. Bon e metastases analysis results in consistently lower estimates of fract ional involvement in bone compared with the conventional region-of-int erest drawing or visual estimation method. This is due to the apparent broadening of objects at and below the limits of resolution of the ga mma camera. Conclusion: image segmentation reduces the delineation and quantitation time of lesions by at least two compared with manual reg ion-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 variabi lity of the user. This method preserves the diagnostic skills of the n uclear medicine physician to select which bony structures contain lesi ons, yet combines it with an objective delineation of the lesion.