A method for size estimation for small objects and its application in brachytherapy seed identification

This study compared the performance of four different object size estimation methods using numerically created experimental images with features relevant to the practice of prostate brachytherapy. The four methods are: 1) pixel count of objects in the segmented binary mask; 2) half peak thresholding based on the previously created binary mask from method 1 and the original image; 3) gray scale averaging of pixels in the binary mask and the surrounding area; and 4) a point spread function corrected version of method 3. The first method demonstrated a 16% error in object size estimation while the other three methods exhibited average errors near 4%. Methods 3 and 4 gave a more consistent estimation of the size for different image contrasts. The first three methods were also applied on three fluoroscopic images of a prostate phantom with 64 implanted seeds. In combination with a classification algorithm, the seed image number and location were determined. Again, method 3 showed superior performance, because it correctly identified 64 seeds in 2 of 3 fluoroscopic images, while 4 false positives appeared in one of the three images. This study confirms that partial volume effects on the size estimation can be compensated by using a gray scale averaging technique. When applied to the seed image identification process, it will improve the accuracy of seed image detection.