Segmentation process significantly influences the accuracy of 3D surface models derived from cone beam computed tomography

详细信息	查看全文 | 推荐本文 |

• 作者：Zacharias Fourie^a ; ^{z.fourie@dmo.umcg.nl} ; Janalt Damstra^a ; Rutger H. Schepers^b ; Peter O. Gerrits^c ; Yijin Ren^a
• 关键词：Cone beam computed tomography ; Three-dimensional ; Segmentation
• 刊名：European Journal of Radiology
• 出版年：2012
• 期刊代码：93_0720048x
• 类别：med
• 出版时间：April, 2012
• 卷：81
• 期：4
• 页码：e524-e530
• 文件大小：724 K

摘要

Aims

To assess the accuracy of surface models derived from 3D cone beam computed tomography (CBCT) with two different segmentation protocols.

Materials and methods

Seven fresh-frozen cadaver heads were used. There was no conflict of interests in this study. CBCT scans were made of the heads and 3D surface models were created of the mandible using two different segmentation protocols. The one series of 3D models was segmented by a commercial software company, while the other series was done by an experienced 3D clinician. The heads were then macerated following a standard process. A high resolution laser surface scanner was used to make a 3D model of the macerated mandibles, which acted as the reference 3D model or 鈥済old standard鈥? The 3D models generated from the two rendering protocols were compared with the 鈥済old standard鈥?using a point-based rigid registration algorithm to superimpose the three 3D models. The linear difference at 25 anatomic and cephalometric landmarks between the laser surface scan and the 3D models generate from the two rendering protocols was measured repeatedly in two sessions with one week interval.

Results

The agreement between the repeated measurement was excellent (ICC = 0.923-1.000). The mean deviation from the gold standard by the 3D models generated from the CS group was 0.330 mm 卤 0.427, while the mean deviation from the Clinician's rendering was 0.763 mm 卤 0.392. The surface models segmented by both CS and DS protocols tend to be larger than those of the reference models. In the DS group, the biggest mean differences with the LSS models were found at the points ConLatR (CI: 0.83-1.23), ConMedR (CI: 鈭?.16 to 2.25), CoLatL (CI: 鈭?.68 to 2.23), Spine (CI: 1.19-2.28), ConAntL (CI: 0.84-1.69), ConSupR (CI: 鈭?.12 to 1.47) and RetMolR (CI: 0.84-1.80).

Conclusion

The Commercially segmented models resembled the reality more closely than the Doctor's segmented models. If 3D models are needed for surgical drilling guides or surgical planning which requires high precision, the additional cost of the commercial segmentation services seem to be justified to produce a more accurate surface models.