UCSF

The purpose of this study was to create a semi-automatic segmentation of the prostate for an accurate estimation of prostate volume and modeling of longitudinal changes in patient data. Segmentation algorithms are available based on axial T2-weighted imaging, but decrease in accuracy in abnormal or treated data, and can be expensive. In this work, we segment based on a fat fraction map (FF map), due to the smaller change in image intensities after treatment and large contrast between the prostate and surrounding fatty tissue. The algorithm consists of five parts: a global filtering of the image; region growth from a predetermined seed point; spline interpolation of the region-enclosing polygon; a comparison of distances between boundary points on adjacent slices; and backward mapping to an upsampled plane with a final volume extrusion. Image artifact is removed by morphological opening after the initial region growth, and by removing boundary points with a shortest distance to an adjacent slice boundary greater than a predetermined threshold. FF maps emphasizing low resolution in favor of high contrast (FF_hc maps), and FF maps emphasizing low contrast in favor of high resolution (FF_hr maps) using different acquisition parameters were tested for accuracy against a manual segmentation drawn on a T2-weighted image in fourteen patients receiving a multiparametric MR exam for confirmed or suspected prostate cancer. There was no significant difference in the volumes recorded from the trial data for the semi-automatic segmentation of the FF_hr map and manual segmentation of the axial T2-weighted image (n=14, p<0.84, paired students t-test), but FF_hr map segmentation had a trend to underestimate the size of the prostate gland. The FF_hc map segmentation volumes were significantly different from those of the axial T2-weighted images (n=14, p<0.03, paired student's t-test). The segmentations of FF_hc maps showed psoas muscle invasion into the prostate region as the most common artifact due to the anatomic proximity and signal similarity, and both FF map types showed ambiguity as the prostate base abuts the bladder, leading to both over- and underestimation. This quick segmentation using FF_hr maps creates accurate prostate volume estimations and should be pursued.