Automated implant segmentation in cone-beam CT using edge detection and particle counting

To develop a fully automated, accurate and robust segmentation technique for dental implants on cone-beam CT (CBCT) images.

A head-size cylindrical polymethyl methacrylate phantom was used, containing titanium rods of 5.15 mm diameter. The phantom was scanned on 17 CBCT devices, using a total of 39 exposure protocols. Images were manually thresholded to verify the applicability of adaptive thresholding and to determine a minimum threshold value \(({T}_{\mathrm{min}})\). A three-step automatic segmentation technique was developed. Firstly, images were pre-thresholded using \({T}_{\mathrm{min}}\). Next, edge enhancement was performed by filtering the image with a Sobel operator. The filtered image was thresholded using an iteratively determined fixed threshold \(({T}_{\mathrm{edge}})\) and converted to binary. Finally, a particle counting method was used to delineate the rods. The segmented area of the titanium rods was compared to the actual area, which was corrected for phantom tilting.

Manual thresholding resulted in large variation in threshold values between CBCTs. After applying the edge-enhancing filter, a stable \({T}_{\mathrm{edge}}\) value of 7.5 % was found. Particle counting successfully detected the rods for all but one device. Deviations between the segmented and real area ranged between \(-\)2.7 and +\(14.4\,\hbox {mm}^{2}\) with an average absolute error of \(2.8\,\hbox {mm}^{2}\). Considering the diameter of the segmented area, submillimeter accuracy was seen for all but two data sets.

A segmentation technique was defined which can be applied to CBCT data for an accurate and fully automatic delineation of titanium rods. The technique was validated in vitro and will be further tested and refined on patient data.