Now Reading
Accuracy of auto-identification of the posteroanterior cephalometric landmarks using cascade convolution neural network algorithm and cephalometric images of different quality from nationwide multiple centers

Accuracy of auto-identification of the posteroanterior cephalometric landmarks using cascade convolution neural network algorithm and cephalometric images of different quality from nationwide multiple centers



American Journal of Orthodontics and Dentofacial Orthopedics, 2022-04-01, Volume 161, Issue 4, Pages e361-e371, Copyright © 2021 American Association of Orthodontists


Introduction

The purpose of this study was to evaluate the accuracy of auto-identification of the posteroanterior (PA) cephalometric landmarks using the cascade convolution neural network (CNN) algorithm and PA cephalogram images of a different quality from nationwide multiple centers nationwide.

Methods

Of the 2798 PA cephalograms from 9 university hospitals, 2418 images (2075 training set and 343 validation set) were used to train the CNN algorithm for auto-identification of 16 PA cephalometric landmarks. Subsequently, 99 pretreatment images from the remaining 380 test set images were used to evaluate the accuracy of auto-identification of the CNN algorithm by comparing with the identification by a human examiner (gold standard) using V-Ceph 8.0 (Ostem, Seoul, South Korea). Pretreatment images were used to eliminate the effects of orthodontic bracket, tube and wire, surgical plate, and surgical screws. Paired t test was performed to compare the x- and y-coordinates of each landmark. The point-to-point error and the successful detection rate (range, within 2.0 mm) were calculated.

Results

The number of landmarks without a significant difference between the location identified by the human examiner and by auto-identification by the CNN algorithm were 8 on the x-coordinate and 5 on the y-coordinate, respectively. The mean point-to-point error was 1.52 mm. The low point-to-point error (<1.0 mm) was observed at the left and right antegonion (0.96 mm and 0.99 mm, respectively) and the high point-to-point error (>2.0 mm) was observed at the maxillary right first molar root apex (2.18 mm). The mean successful detection rate of auto-identification was 83.3%.

Conclusions

Cascade CNN algorithm for auto-identification of PA cephalometric landmarks showed a possibility of an effective alternative to manual identification.

Highlights

  • Cascade CNN algorithm showed clinically acceptable accuracy for auto-identification of PA cephalometric landmarks.

  • This study was the first to investigate the accuracy of auto-identification of PA cephalometric landmarks using the cascade CNN algorithm.

  • This study is the first to use PA cephalogram images of different qualities from 9 nationwide multicenters.

  • Cascaded CNN algorithm for auto-identification of PA cephalometric landmarks showed a possibility of an effective alternative to manual identification.

The artificial intelligence (AI) is defined as the ability of machine to mimic intelligent human behavior for problem solving, object recognition, and decision-making. , The substructures of AI are machine learning and deep learning. Machine learning is an algorithm that allows computers to learn on their own through data without explicit programming using statistical techniques, and so on. Of the many algorithms of machine learning, there are algorithms called artificial neural networks (ANNs) based on perceptron. When the network becomes deeper, it is called deep learning. Deep learning is the application of ANNs containing multiple hidden layers. The representative algorithms of deep learning are convolution neural network (CNN) and recurrent neural network. ,

AI technology has been applied to various medical fields including diagnostic imaging, genetics, electrical diagnosis, monitoring, disability evaluation, and mass screening. In the field of dentistry, AI technology is usually applied for the diagnosis of caries, oral cancer, and osteomyelitis and to evaluate the necessity of tooth extraction using radiographic images. In orthodontics, AI has been usually applied to automated landmark identification, diagnosis, and data mining. Cephalometric analysis is essential to establish an accurate diagnosis and treatment plan, and its first step is accurate identification of the cephalometric landmarks. However, this traditional manual procedure is laborious and time-consuming; furthermore, the quality of diagnosis using cephalometric analysis depends on the accuracy and reliability of landmark identification. ,

Therefore, numerous studies have attempted to evaluate auto-identification of lateral cephalometric landmarks using AI with machine learning or deep learning. , Arık et al reported 75.3% of successful detection rate (SDR) within 2.0 mm range using a CNN for binary classification. Lindner et al reported 74.8% of SDR within 2.0 mm range using the algorithm based on random forest. Hwang et al reported a point-to-point error of 1.46 mm using deep learning method (YOLOv3).

However, these AI studies , had several limitations as follows: (1) these studies reported the accuracy of auto-identification of the lateral cephalometric landmarks only. Lateral cephalometric analysis can define the anteroposterior and vertical problems. However, because cephalometric analysis of the dentoalveolar and/or skeletal asymmetry and transverse discrepancy is also important, it is necessary to develop an AI algorithm for auto-identification of the posteroanterior (PA) cephalometric landmarks. (2) These studies used the algorithms based on machine learning or simple deep learning such as one-stage CNN. However, it is necessary to develop an AI algorithm from simple machine learning to complex deep learning for detection of the exact location of landmarks from cephalograms of large sizes. (3) The data set of these studies were collected from only 1 or 2 centers. Therefore, it is necessary to train the AI algorithm using cephalograms obtained from multiple centers for general use.

Furthermore, to the best of the authors’ knowledge, there is only one study on accuracy of auto-identification of the PA cephalometric landmarks using AI until now. Muraev et al reported that the value of the mean absolute error for the points placed by the ANN was similar to that of the expert group (2.87 mm vs 2.47 mm). However, that study used ANN and only 300 PA cephalograms as training set and 30 PA cephalograms as test set form 1 center. Therefore, the purpose of our study was to investigate the accuracy of auto-identification of the PA cephalometric landmarks using cascade CNN algorithm and PA cephalogram images of different quality from nationwide 9 university hospitals.

Material and methods

A total of 2798 PA cephalograms were collected from the Department of Orthodontics of 9 university hospitals, including Seoul National University Dental Hospital (SNUDH; n = 1489), Kyung Hee University Dental Hospital (n = 615), Kyungpook National University Dental Hospital (KNUDH; n = 180), Asan Medical Center (n = 137), Ajou University Dental Hospital (n = 110), Korea University Dental Hospital (n = 85), Chonnam National University Dental Hospital (CNUDH; n = 79), Wonkwang University Dental Hospital (n = 75), and Ewha University Medical Center (EUMC; n = 28) ( Table I ). This study was reviewed and approved by the Institutional Review Board Committee of 9 multicenters; including SNUDH (ERI18002), Kyung Hee University Dental Hospital (D19-007-003), KNUDH (KNUDH-2019-03-02-00), Asan Medical Center (2019-0927), Ajou University Dental Hospital (AJIRB-MED-MDB-19-039), Korea University Dental Hospital (2019AN0166), CNUDH (CNUDH-2019-004), Wonkwang University Dental Hospital (WKDIRB202010-06) and EUMC (EUMC 2019-04-017-009). The requirement for patient consent was waived by the Institutional Review Board Committee of each center.

Table I
Composition of the dataset of PA cephalogram images from nationwide 9 university hospitals
Hospitals Training set Validation set Test set Sum
Pretreatment Presurgical Postsurgical Posttreatment
SNUDH 1290 100 25 23 29 22 1489
KHUDH 393 100 24 43 41 14 615
KNUDH 120 30 9 6 10 5 180
AMC 74 30 10 7 8 8 137
AUDH 70 20 3 5 5 7 110
KUDH 47 19 8 5 6 0 85
CNUDH 34 20 11 2 6 6 79
WUDH 27 20 9 10 9 0 75
EUMC 20 4 0 1 3 0 28
Sum 2075 343 99 102 117 62 2798
KHUDH, Kyung Hee University Dental Hospital; AMC, Asan Medical Center; AUDH, Ajou University Dental Hospital; KUDH, Korea University Dental Hospital; WUDH, Wonkwang University Dental Hospital.

Only pretreatment images (n = 99) were selected to investigate the accuracy of auto-identification.

Inclusion criteria were as follows: (1) patients who were Korean adults after completion of facial growth, (2) patients who had treatment plans for orthognathic surgery or had already undergone orthognathic surgery between 2013 and 2020, and (3) patients who had permanent dentition. Exclusion criterion was patients who had dentofacial traumas, craniofacial syndromes, or systemic diseases. All data sets were constructed without restrictions on gender.

All PA cephalograms were converted into gray scale images with 2k × 2k pixel and 8-bit depth, stored in Digital Imaging and Communications in Medicine file format. Considering the ratio of the original image size, all PA cephalogram images were resized to 700 × 1000 pixel, and pixel normalization was performed by dividing it by 255.0 to have pixel values in the range 0-1.

All the 2798 PA cephalogram images were randomly allocated into training set (n = 2075 images), internal validation set (n = 343 images), and test set (n = 380 images). The training and internal validation sets were used regardless of the patient's treatment status. However, in the test set, only pretreatment images (n = 99 images) were selected to evaluate the accuracy of auto-identification for eliminating the effects of orthodontic bracket, tube and wire, surgical plate, and surgical screws.

We used a fully automated landmark identification algorithm with a cascade CNN that consisted of the region of interest (ROI) detection stage and the landmark prediction stage, to improve the prediction accuracy and reduce false-positive regions ( Fig 1 ) .

Cascade CNN algorithm used in this study. Stage 1, the ROI detection to propose the area of interest; stage 2, the landmark prediction to find the exact location of landmarks.
Fig 1
Cascade CNN algorithm used in this study. Stage 1, the ROI detection to propose the area of interest; stage 2, the landmark prediction to find the exact location of landmarks.

First, candidate ROI of various sizes were trained by the ROI detection stage. For stronger ROI detection, the complexity of the regions surrounding each landmark should be considered. RetinaNet was used to detect ROI. The ROI patches with the center of the landmark were extracted with coordinates (Tx, Ty). For identifying each landmark, 2 sizes of ROI including 256 x 256 pixel and 512 x 512 pixel were evaluated by a single examiner (S.M.G), who had finished 3-year orthodontic resident program and had obtained board certification for orthodontic specialty ( Fig 2 ).

Examples of ROI size in the PA cephalometric radiograph. Red , the crista galli and the maxillary central incisor crown requiring a small ROI with 256 × 256 pixel size; blue , the right side of the frontozygomatic point requiring a wide ROI with 512 × 512 pixel size.
Fig 2
Examples of ROI size in the PA cephalometric radiograph. Red , the crista galli and the maxillary central incisor crown requiring a small ROI with 256 × 256 pixel size; blue , the right side of the frontozygomatic point requiring a wide ROI with 512 × 512 pixel size.

Second, the U-Net was used to find the exact locations within the ROI patches from the first stage. The center of ROI patches was represented as (|Tx-Dx |, |Ty-Dy |) and the ROI detection’s mean distance errors (Dx, Dy) were extracted. The circular segmentation labels with the diameter d were generated at the center of ROI and the most appropriate d was experimentally determined as 50 pixels.

Various augmentation methods such as gaussian noise, random brightness, blurring, random contract, flip, and random rotation were used to train deep learning model. Setting learning rate to 0.0001, the adam optimizer was used. And the loss function and the model performance were calculating by the dice similarity coefficient.

The landmarks used in this study were adopted from previous studies on PA cephalometric landmarks ( Fig 3 ; Table II ). ,

Landmarks in PA cephalometric radiograph. Cg , crista galli; R FZP , the right frontozygomatic point; L FZP , the left frontozygomatic point; R J , the right jugal point; L J , the left jugal point; R Ag , the right antegonion; L Ag , the left antegonion; Me , menton; R U6C , the maxillary right first molar crown; L U6C , the maxillary left first molar crown; R U6R , the maxillary right first molar root apex; L U6R , the maxillary left first molar root apex; R U1C , the maxillary right central incisor crown; L U1C , the maxillary left central incisor crown; R U1R , the maxillary right central incisor root apex; L U1R , the maxillary left central incisor root apex.
Fig 3
Landmarks in PA cephalometric radiograph. Cg , crista galli; R FZP , the right frontozygomatic point; L FZP , the left frontozygomatic point; R J , the right jugal point; L J , the left jugal point; R Ag , the right antegonion; L Ag , the left antegonion; Me , menton; R U6C , the maxillary right first molar crown; L U6C , the maxillary left first molar crown; R U6R , the maxillary right first molar root apex; L U6R , the maxillary left first molar root apex; R U1C , the maxillary right central incisor crown; L U1C , the maxillary left central incisor crown; R U1R , the maxillary right central incisor root apex; L U1R , the maxillary left central incisor root apex.

Table II
Definition of landmarks
Landmarks Definition
Midline landmarks
Cg (Crista galli) The middle point of crista galli
Me (Menton) The most inferior point of symphysis of mandible
Bilateral landmarks
FZP (Frontozygomatic point) The intersection of the frontozygomatic suture and the inner rim of the orbit
J (Jugal point) The intersection of the tuberosity of maxilla and zygomatic buttress
Ag (Antegonion) The antegonial notch at the lateral inferior margin of the antegonial protuberances
U6C The most lateral point of maxillary first molar crown
U6R The most lateral point of maxillary first molar root apex
U1C The middle point of incisal edge of maxillary central incisor
U1R The root apex of maxillary central incisor

The orthodontists in SNUDH manually traced 16 PA cephalometric landmarks using V-Ceph 8.0 (Ostem, Seoul, South Korea). The first training was performed using this training set of SNUDH (n = 1290 images).

This trained algorithm identified the PA cephalometric landmarks using the training set of the remaining other 8 hospitals (n = 785 images). Subsequently, these landmarks were checked and modified by orthodontists of the 8 hospitals, who were pre-educated on the same definition of landmarks. The second training was conducted with the modified landmarks of the 8 hospitals.

The internal validation set (n = 343 images) was used to verify the hyperparameter of model selection.

After the internal validation, the finally selected model with cascade CNN algorithm automatically identified 16 PA cephalometric landmarks using 99 pretreatment test set images.

A single examiner (S.M.G), who has finished 3-year orthodontic resident program and has obtained an orthodontic specialty board, was selected as gold standard. After reading the definition of 16 PA cephalometric landmarks, all landmarks were manually pointed on 99 pretreatment test set images by using a mouse-controlled cursor linked to the V-Ceph 8.0 (Ostem) ( Table II ). The landmark positions on each image were expressed in the x- and y-coordinates (horizontal plane and vertical plane, respectively). Human examiner manually identified 16 PA cephalometric landmarks on 99 pretreatment PA cephalograms twice with a 2-week interval. Subsequently, the mean values of each coordinates (x, y) were set as the gold standard of the PA cephalometric landmarks.

The mean point-to-point error between repeated trials by human examiner is 0.73 ± 0.49 mm ( Table III ). Because the range of intraclass correlation coefficients of the x- and y-coordinates of the 16 PA cephalometric landmarks was 0.993-1.000, it indicated excellent intraobserver reliability. Therefore, the mean values of the first and second sets of identification were used for further statistical analysis.

Table III
The point-to-point error and intraobserver reliability of human examiner (gold standard)
Variables Point-to-point error (mm)
(n = 99 pretreatment PA cephalograms)
Intraclass correlation coefficient value
(n = 99 pretreatment PA cephalograms)
Mean ± SD x-coordinates y-coordinates
Crista galli 0.72 ± 0.57 0.999 ∗∗∗ 0.997 ∗∗∗
Menton 1.04 ± 0.57 0.993 ∗∗∗ 1.000 ∗∗∗
Frontozygomatic point
right 0.92 ± 0.62 0.993 ∗∗∗ 0.997 ∗∗∗
left 0.83 ± 0.50 0.999 ∗∗∗ 0.997 ∗∗∗
Jugal point
right 0.72 ± 0.47 0.996 ∗∗∗ 0.998 ∗∗∗
left 0.73 ± 0.52 0.999 ∗∗∗ 0.999 ∗∗∗
Antegonion
right 0.65 ± 0.38 0.997 ∗∗∗ 0.999 ∗∗∗
left 0.65 ± 0.51 0.999 ∗∗∗ 0.999 ∗∗∗
U6C
right 0.62 ± 0.43 0.998 ∗∗∗ 0.999 ∗∗∗
left 0.69 ± 0.58 0.999 ∗∗∗ 1.000 ∗∗∗
U6R
right 0.82 ± 0.61 0.996 ∗∗∗ 0.999 ∗∗∗
left 0.96 ± 0.60 0.999 ∗∗∗ 0.999 ∗∗∗
U1C
right 0.56 ± 0.34 0.999 ∗∗∗ 0.999 ∗∗∗
left 0.51 ± 0.34 0.999 ∗∗∗ 0.999 ∗∗∗
U1R
right 0.64 ± 0.42 0.999 ∗∗∗ 0.999 ∗∗∗
left 0.65 ± 0.46 0.998 ∗∗∗ 0.998 ∗∗∗
Average 0.73 ± 0.49
Note. The intraclass reliability test was performed by 2-way random effects model.
U6C , maxillary first molar crown; U6R , maxillary first molar root apex; U1C , maxillary central incisor crown; U1R , maxillary central incisor root apex.

∗∗∗ P <0.001.

The degree of the point-to-point error of the PA cephalometric landmarks between auto-identification and human examiner (gold standard) and the distribution of SDR in the auto-identification were calculated. The point-to-point error indicates the absolute distance difference between auto-identification and manual identification by human examiner (gold standard), calculated in millimeter scales converted through pixel spacing. The accuracy of ours proposed model was expressed using the Euclidean distance between the ground truth (Tx, Ty) and the predicted landmarks (Px, Py) from our proposed model. Because this distance can be considered as pixel distance, the point-to-point error was obtained by multiplying the pixel spacing of the radiograph, where pixel spacing ranges from 0.1 mm to 0.13 mm. The SDR indicates what percentage of PA cephalometric landmark was identified within a precision range. , ,

Statistical analysis

Paired t test and intraclass reliability test were performed using SPSS software (version 12.0; SPSS Inc, Chicago, Ill). The P value <0.05 was considered to be statistically significant.

Results

To visualize the difference between PA cephalometric landmarks identified by human examiner (gold standard) and auto-identification, 16 landmarks in each PA cephalogram image were superimposed ( Fig 4 ). The manual identification by human examiner (gold standard) was shown in red and the auto-identification by cascade CNN algorithm was show in green.

Examples of superimposition of the identified PA cephalometric landmarks. Green , auto-identification by cascade CNN algorithm; red , manual identification by human examiner (gold standard).
Fig 4
Examples of superimposition of the identified PA cephalometric landmarks. Green , auto-identification by cascade CNN algorithm; red , manual identification by human examiner (gold standard).

There were no significant differences in the pixel location of 8 landmarks in the x-coordinate (crista galli, right frontozygomatic point, left antegonion, maxillary right first molar crown, maxillary right central incisor crown, maxillary left central incisor crown, maxillary right central incisor root apex, maxillary left central incisor root apex, all P >0.05), and 5 landmarks in the y-coordinate (crista galli, left antegonion, maxillary left first molar crown, maxillary right central incisor crown, maxillary left central incisor crown, all P >0.05) ( Table IV ).

Table IV
Comparison of the x- and y-coordinates of the PA cephalometric landmarks identified by human examiner (gold standard) and auto-identification
Variables The x-coordinates of the PA cephalometric landmarks (pixel) The y-coordinates of the PA cephalometric landmarks (pixel)
Human examiner (gold standard)
(n = 99 pretreatment PA cephalograms)
Auto-identification (n = 99 pretreatment PA cephalograms) △human examiner-Auto-identification P value Human examiner (gold standard)
(n = 99 pretreatment PA cephalograms)
Auto-identification (n = 99 pretreatment PA cephalograms) △human examiner-Auto-identification P value
Mean ± SD Mean ± SD Mean ± SD Mean ± SD Mean ± SD Mean ± SD
Crista galli 920.8 ± 113.4 921.5 ± 113.8 −0.7 ± 4.05 0.095 927.2 ± 182.7 931.6 ± 184.6 −4.4 ± 19.8 0.529
Menton 938.1 ± 117.9 925.2 ± 117.4 12.9 ± 14.1 0.000 ∗∗∗ 1965.6 ± 315.4 1968.4 ± 317.2 −2.7 ± 6.7 0.000 ∗∗∗
Frontozygomatic point
right 529.4 ± 91.1 528.9 ± 93.5 0.5 ± 9.8 0.647 943.1 ± 185.6 938.2 ± 179.2 4.9 ± 15.8 0.003 ∗∗
left 1313.1 ± 159.8 1308.8 ± 158.8 4.3 ± 8.4 0.000 ∗∗∗ 943.3 ± 184.8 937.5 ± 178.2 5.8 ± 16.4 0.001 ∗∗
Jugal point
right 634.8 ± 93.6 637.3 ± 93.6 6.5 ± 10.0 0.000 ∗∗∗ 1401.1 ± 232.9 1394.4 ± 230.3 6.7 ± 12.8 0.000 ∗∗∗
left 1200.3 ± 146.6 1202.0 ± 148.1 −1.7 ± 7.4 0.022 1399.0 ± 234.0 1396.0 ± 232.5 3.1 ± 11.7 0.010
Antegonion
right 543.6 ± 94.3 539.6 ± 95.1 4.1 ± 6.5 0.000 ∗∗∗ 1747.1 ± 276.6 1743.2 ± 275.6 3.9 ± 6.8 0.000 ∗∗∗
left 1301.7 ± 159.5 1302.0 ± 160.1 −0.4 ± 6.5 0.587 1741.8 ± 282.5 1743.4 ± 281.0 −1.5 ± 12.1 0.213
U6C
right 673.9 ± 94.4 672.0 ± 94.6 2.0 ± 10.5 0.061 1554.2 ± 251.4 1550.1 ± 251.8 4.1 ± 11.1 0.000 ∗∗∗
left 1173.0 ± 143.7 1170.2 ± 144.4 2.8 ± 8.6 0.001 ∗∗ 1633.5 ± 266.1 1633.5 ± 267.3 0.0 ± 7.6 0.984
U6R
right 716.6 ± 99.4 714.6 ± 97.2 2.0 ± 14.3 0.006 ∗∗ 1440.5 ± 236.9 1432.6 ± 235.7 7.9 ± 12.9 0.000 ∗∗∗
left 1131.8 ± 138.3 1127.0 ± 136.9 4.8 ± 10.6 0.000 ∗∗∗ 1464.2 ± 243.6 1459.5 ± 242.0 4.7 ± 10.9 0.000 ∗∗∗
U1C
right 885.9 ± 109.8 886.9 ± 109.9 −1.0 ± 7.4 0.185 1633.5 ± 265.4 1632.1 ± 270.3 1.4 ± 21.6 0.531
left 960.1 ± 117.8 958.4 ± 118.0 1.7 ± 10.9 0.126 1552.7 ± 253.6 1552.5 ± 254.2 0.2 ± 9.4 0.836
U1R
right 897.5 ± 111.7 896.4 ± 110.7 1.1 ± 7.9 0.173 1465.2 ± 242.9 1460.6 ± 242.1 4.7 ± 11.8 0.000 ∗∗∗
left 948.3 ± 116.5 947.1 ± 115.6 1.2 ± 6.1 0.051 1442.3 ± 239.1 1435.3 ± 236.4 7.0 ± 12.5 0.000 ∗∗∗
Note. Pixel size, the Euclidean distance between the ground truth and the predicted landmarks from the proposed model. The paired t test were performed.
U6C , maxillary first molar crown; U6R , maxillary first molar root apex; U1C , maxillary central incisor crown; U1R , maxillary central incisor root apex.

P < 0.05.

∗∗ P < 0.01.

∗∗∗ P < 0.001.

The mean value of the point-to-point error was 1.52 mm between identification by human examiner (gold standard) and auto-identification ( Table V ; Fig 5 ).

Table V
The point-to-point error between identification by human examiner (gold standard) and auto-identification
Variables Point-to-point error (mm)
(n = 99 pretreatment PA cephalograms)
Mean ± SD
Crista galli 1.89 ± 1.61
Menton 1.99 ± 1.44
Frontozygomatic point
right 1.83 ± 1.42
left 1.96 ± 1.33
Jugal point
right 1.71 ± 1.45
left 1.30 ± 1.09
Antegonion
right 0.99 ± 0.87
left 0.96 ± 1.35
U6C
right 1.59 ± 1.06
left 1.33 ± 0.83
U6R
right 2.18 ± 1.23
left 1.93 ± 1.08
U1C
right 1.05 ± 0.84
left 1.12 ± 1.16
U1R
right 1.53 ± 0.95
left 1.37 ± 0.85
Average 1.52 ± 1.13
U6C , maxillary first molar crown; U6R , maxillary first molar root apex; U1C , maxillary central incisor crown; U1R , maxillary central incisor root apex.

Mean values of the point-to-point errors of PA cephalometric landmarks between identification by human examiner and auto-identification. Red , clinically acceptable error within 2.0 mm; green , mean error in this study. U6C , maxillary first molar crown; U6R , maxillary first molar root apex; U1C , maxillary central incisor crown; U1R , maxillary central incisor root apex.
Fig 5
Mean values of the point-to-point errors of PA cephalometric landmarks between identification by human examiner and auto-identification. Red , clinically acceptable error within 2.0 mm; green , mean error in this study. U6C , maxillary first molar crown; U6R , maxillary first molar root apex; U1C , maxillary central incisor crown; U1R , maxillary central incisor root apex.

In terms of the point-to-point error between human examiner and auto-identification, the lower value of point-to-point error (<1.0 mm) was observed at the left and right antegonion (0.96 mm and 0.99 mm, respectively). On the other hand, the higher value of point-to-point error (>2.0 mm) was observed at the maxillary right first molar root apex (2.18 mm).

Compared with identification by human examiner (gold standard), the average of SDRs of the auto-identification were 47.9% within range of 1.0 mm, 68.2% within range of 1.5 mm, 83.3% within range of 2.0 mm, 90.6% within range of 2.5 mm, 94.9% within range of 3.0 mm, and 98.0% within range of 4.0 mm ( Table VI ). More than 80% of SDR was observed from range of 2.0 mm to range of 4.0 mm in auto-identification.

Table VI
Distribution of the SDR in the auto-identification
Variables SDR (%)
(n = 99 pretreatment PA cephalograms)
<1.0 mm <1.5 mm <2.0 mm <2.5 mm <3.0 mm <4.0 mm
Crista galli 40.4 59.6 74.7 83.8 92.9 94.9
Menton 35.4 57.6 72.7 82.8 87.9 94.9
Frontozygomatic point
right 39.4 64.6 77.8 83.8 87.9 96.0
left 31.3 50.5 70.7 86.9 88.9 94.9
Jugal point
right 47.5 60.6 72.7 86.9 94.9 96.0
left 61.6 76.8 89.9 91.9 96.0 100
Antegonion
right 71.7 87.9 96.0 98.0 99.0 99.0
left 75.8 90.9 97.0 99.0 100 100
U6C
right 43.4 67.7 81.8 88.9 96.0 99.0
left 53.5 77.8 89.9 93.9 96.0 99.0
U6R
right 21.2 43.4 65.7 78.8 87.9 97.0
left 25.3 46.5 76.8 90.9 97.0 99.0
U1C
right 70.7 82.8 96.0 98.0 100 100
left 65.7 82.8 92.9 97.0 99.0 100
U1R
right 40.4 70.7 87.9 92.9 97.0 99.0
left 43.4 70.7 89.9 96.0 99.0 100
Average 47.9 68.2 83.3 90.6 94.9 98.0
U6C , maxillary first molar crown; U6R , maxillary first molar root apex; U1C , maxillary central incisor crown; U1R , maxillary central incisor root apex.

Within a range of 2.0 mm, the auto-identification showed a high SDRs (more than 90%) and they were observed at the left antegonion (97.0%), the right antegonion (96.0%), the maxillary right central incisor crown (96.0%), and the maxillary left central incisor crown (92.9%), and the low SDR (less than 70%) was observed at the maxillary right first molar root apex (65.7%).

You're Reading a Preview

Become a DentistryKey membership for Full access and enjoy Unlimited articles

Become membership

If you are a member. Log in here