You are on page 1of 7

BASIC RESEARCH – TECHNOLOGY

Frank C. Setzer, DMD, PhD,


Artificial Intelligence for the MS,* Katherine J. Shi, DMD,†
Zhiyang Zhang, BS,‡ Hao Yan,
Computer-aided Detection of PhD,‡ Hyunsoo Yoon, PhD,‡
Mel Mupparapu, DMD, MDS,
Periapical Lesions in Cone- DipABOMR,x and Jing Li, PhD‡

beam Computed Tomographic


Images

ABSTRACT
SIGNIFICANCE
Introduction: The aim of this study was to use a Deep Learning (DL) algorithm for the
automated segmentation of cone-beam computed tomographic (CBCT) images and the Artificial intelligence has been
detection of periapical lesions. Methods: Limited field of view CBCT volumes (n 5 20) widely used in biomedical
containing 61 roots with and without lesions were segmented clinician dependent versus using image analysis in other medical
the DL approach based on a U-Net architecture. Segmentation labeled each voxel as 1 of 5 fields. The application of Deep
categories: “lesion” (periapical lesion), “tooth structure,” “bone,” “restorative materials,” and Learning algorithms to identify
“background.” Repeated splits of all images into a training set and a validation set based on 5- periapical lesions in CBCT
fold cross validation were performed using Deep Learning segmentation (DLS), and the results volumes may enhance
were averaged. DLS versus clinical-dependent segmentation was assessed by dichotomized dentists’ diagnostic efforts.
lesion detection accuracy evaluating sensitivity, specificity, positive predictive value, negative
predictive value, and voxel-matching accuracy using the DICE index for each of the 5 labels.
Results: DLS lesion detection accuracy was 0.93 with specificity of 0.88, positive predictive
value of 0.87, and negative predictive value of 0.93. The overall cumulative DICE indexes for the
individual labels were lesion 5 0.52, tooth structure 5 0.74, bone 5 0.78, restorative materials
5 0.58, and background 5 0.95. The cumulative DICE index for all actual true lesions was
0.67. Conclusions: This DL algorithm trained in a limited CBCT environment showed
excellent results in lesion detection accuracy. Overall voxel-matching accuracy may be
benefited by enhanced versions of artificial intelligence. (J Endod 2020;-:1–7.)

KEY WORDS
Artificial intelligence; cone-beam computed tomography; Deep Learning; digital imaging/
radiology; periapical lesion; U-Net

Apical periodontitis is the primary disease in endodontics, comprising about 75% of radiolucent jaw From the Departments of *Endodontics
and xOral Medicine, School of Dental
lesions1. It is highly prevalent, affecting 33%–62% of the population in age groups ranging from 20 to .60 Medicine, University of Pennsylvania,
years2. The principal diagnostic tool for the detection of apical periodontitis is intraoral dental radiography Philadelphia, Pennsylvania; †Private
by means of detecting periapical lesions (PLs). This is associated with several challenges and limitations. Practice, University of Pennsylvania,
Traditional 2-dimensional (2D) radiographs depict a summation of all content in the X-ray pathway and its Philadelphia, Pennsylvania; and ‡School
of Computing, Informatics, and Decision
attenuation within the structures. This content-rich environment of the oral cavity and jaws, including
Systems Engineering, Arizona State
different structures, tissues, and restorative materials, greatly increases the chance that lesions, University, Tempe, Arizona
particularly during early stages, will be masked3. Moreover, detection and a differential diagnosis rely on
Address requests for reprints to Dr Frank
observer interpretation, which is known to be inconsistent4,5. C. Setzer, School of Dental Medicine,
Three-dimensional (3D) imaging technology in the form of cone-beam computed tomographic University of Pennsylvania, 240 S 40th
(CBCT) imaging is now common in clinical practice with approximately 80% utilization among Street, Philadelphia, PA 19104.
endodontists for diagnosis and treatment planning6. Compared with 2D periapical radiography, CBCT E-mail address: fsetzer@upenn.edu
0099-2399/$ - see front matter
imaging has increased the accuracy of detecting periapical lesions. A meta-analysis reported a pooled
sensitivity and specificity of 0.70 and 0.78 for digital radiography compared with 0.95 and 0.88 for CBCT Copyright © 2020 American Association
of Endodontists.
imaging, respectively7. However, in clinically relevant situations, such as smaller lesions on root-filled
https://doi.org/10.1016/
teeth, CBCT accuracy is greatly reduced (sensitivity 5 0.63, specificity 5 0.69)8. Moreover, clinician- j.joen.2020.03.025

JOE  Volume -, Number -, - 2020 Artificial Intelligence for Detecting Periapical Lesions 1
dependent interpretation of CBCT imaging still knowing the shape of a lesion, which, moving CBCT analysis and segmentation and included
suffers from low inter- and intraobserver forward, will aid in facilitating a subsequent an oral and maxillofacial radiologist (M.M.), an
agreement. An observer consensus range of differential diagnosis. endodontist (F.C.S.), and a senior graduate
only 0.32–0.49 for PLs was reported9. student completing a radiology honors
Computer-aided detection and program (K.J.S.). Disagreements were
MATERIALS AND METHODS
diagnosis (CAD) has been widely applied to resolved by joint discussion. After complete
biomedical image analysis outside of dentistry. CBCT Selection segmentation and review, the most
CAD systems are computer applications that Institutional approval was obtained representative sagittal region (traversing the
assist in the detection and/or diagnosis of (institutional review board #834136). Existing roots and PL if applicable) was determined,
diseases by providing an unbiased “second deidentified CBCT volumes of 20 patients with including a center slice and 2 additional slices
opinion” to the image interpreter10, aiming at known age and sex were selected based on in each buccal and lingual direction. Image
improving accuracy and reducing time for criteria allowing for training and cross validation analyses was performed using standardized
analysis11. Computer-aided detection (CADe) of the DL algorithm to detect PLs. The monitor settings and lighting conditions.
systems identify the presence and location of inclusion criteria were as follows: limited field of
lesions in biomedical images. Computer-aided view containing several roots with at least 1
diagnosis (CADx) systems characterize lesions periapical lesion per volume, teeth with or CBCT Segmentation and Labeling
(eg, they distinguish between benign and without a history of endodontic treatment, Using the DL Approach
malignant tumors). With the rapid growth of teeth with or without definitive coronal All 20 pairs of deidentified RAW and clinician-
Deep Learning (DL) algorithms in image-based restoration, and no to moderate presence of dependent processed CBCT volumes,
applications, CAD systems can now be trained scatter and/or artifacts. CBCT volumes were including the information on the representative
by DL to provide more advanced capability (ie, normalized to contain the aforementioned region, were forwarded to 3 experts in AI-
the capability of artificial intelligence [AI]) to best parameters. All CBCT images were acquired based biomedical image analysis (J.L., H.Y.,
assist clinicians. using a Morita Veraviewpocs 3DF40 (Morita and Z.Z.) at Arizona State University. By
Few studies investigated CAD systems Mfg, Kyoto, Japan) with a field of view of 40 remote training and discussion, an initial
in dentistry. DL/AI-based CAD systems are mm at a voxel size of 125 mm from patients version of the DL algorithm had been
rare. Existing applications are still insufficient in referred for endodontic treatment. The reasons developed to test the ability of DL in CBCT
improving the objectivity and accuracy of for referral and subsequent treatment were not segmentation and PL detection (referred to as
lesion detection and diagnosis because of recorded for this study. All CBCT volumes Deep Learning segmentation [DLS]). Briefly, a
several limitations. The majority of algorithms were initially reviewed on 24-inch light-emitting multilabel U-Net26 was used as the baseline
are semiautomated and thus require a user diode monitors (resolution 1920 ! 1200 at 60 architecture, focusing on the 5 labels specified
selection of seed points followed by manual Hz) using the proprietary software provided by previously. A detailed description of U-
adjustments. The accuracy of these algorithms the CBCT company (iDixel, Morita Mfg). A PL Net algorithms in general and its application in
relies heavily on user experience. Automated was identified when a disruption of the lamina biomedical image segmentation is provided
algorithms based on thresholding12–14, region dura was detected, and the low-density area elsewhere27. For the 5 selected slices of each
growing15–17, and statistical methods18,19 associated with the radiographic apex was at CBCT image, individual U-Nets with shared
have been published but have not reached least twice the width of the periodontal parameters were trained, respectively. In the
clinical practice. Most algorithms are based on ligament space20–23. last layer of DLS, a 1 ! 1 3D convolution layer
2D single slices of CBCT volumes or was used to integrate information from the 5 U-
panoramic radiographs, which provide only a Clinician-dependent CBCT Nets for segmentation of the center slice
singular view12–19. Because of this deficiency, Segmentation and Labeling (Fig. 2). Because DLS was used to perform
some lesions may be obscured and missed. In Clinician-dependent CBCT volume multilabel segmentation, unequal weights for
addition, radiographic characteristics of segmentation (referred to as clinician the different labels were assigned in the loss
lesions may not fully present, causing segmentation [CS]) was performed using ITK- function throughout training. The label lesion
difficulties for attempts at differential SNAP (open-source) following protocols was attributed 8! the weight of the other
diagnosis17. established in previous studies23–25. From the labels because it is the label with the most
Our short-term goals for the present unprocessed CBCT images (referred to as clinical relevance. Furthermore, after DLS had
study was to develop and validate a DL RAW), 5 standardized categories of image generated a segmented voxel-wise multilabel
platform that segments “lesion” (PL), “tooth content were identified and labeled: lesion (PL), map for each slice, the map was cleaned by
structure,” “bone,” “restorative materials,” and tooth structure, bone, restorative materials implementing automatic rules based on
“background” (5 labels) on CBCT images and (including root filling), and background domain knowledge of anatomic constraints of
to compare its lesion detection accuracy to (everything else). For each label, the process PLs (eg, proximity to roots).
that of experienced clinicians. Our long-term started with a semiautomatic segmentation DLS was then taught to segment the 5
goal is to develop an advanced CADe/CADx algorithm in the Active Contour segmentation labels previously described based on 5-fold
platform empowered by AI to automate the mode on ITK-SNAP, including the stages of cross validation. In detail, 20 CBCT images
segmentation of whole CBCT volumes, detect presegmentation (determination of the region were divided into 5 folds. Four folds (16 CBCT
lesions with high accuracy and efficiency, and of interest), initialization (manual insertion of images) were used for training, and the
ultimately differentiate different types of oral seeds), and evolution (expansion of seeds remaining fold (4 CBCT images) was used for
pathoses with high-precision AIDen (Artificial labeling the area of interest) (Fig. 1A–D). After validation. This process was iterated with every
Intelligence in Dentistry). The segmentation all labels were initially completed, each sagittal, fold used as the validation fold (ie, 5 total splits
algorithm developed for this study coronal, and axial slice was reviewed and of the data). The average DICE index of lesion
automatically leads to the detection of lesions manually revised by 3 reviewers, if necessary. over the 5 iterations was used as the criterion
but also provides the additional capacity of All reviewers were calibrated and trained in to optimize the DL algorithm. DICE is a

2 Setzer et al. JOE  Volume -, Number -, - 2020


FIGURE 1 – CS. (A ) RAW CBCT image in the axial, coronal, and sagittal direction with identification of the region of interest (red box ). (B ) Initialization of segmentation by manual
insertion of seed points after presegmentation with thresholding. (C ) Final segmentation of tooth structure of lower anterior after seed point evolution and manual correction. (D ) Final
segmentation of all labels present in the image. Each color indicates a different label and was separately segmented.

commonly used metric to evaluate and 69% fair, and ,60% as poor. Roots presence of a PL was identified based on the
segmentation accuracy. Please see the next without lesions served as controls. criteria outlined previously, and 32 of 61 roots
section for detailed definition. 2. Voxel-matching accuracy using the (52.5) had no lesions (Table 1). The initial
DICE index calculated for each label: for agreement between the reviewers was 27 of
example for lesion, after our algorithm 29 lesions (93.1%). The 2 discrepancies were
Data Analysis and Statistical generated a segmented image, we compared resolved by joint review and discussion. DLS
Evaluation the set of pixels labeled as lesion by the DLS identified 27 true-positive and 28 true-negative
Repeated splits of all images based on the 5- algorithm and the set of pixels labeled as lesion situations. There were 4 false-positive and 2
fold cross validation were performed using by CS. Then, the DICE index of the lesion was false-negative accounts. For the dichotomized
DLS, and the results were averaged. The computed as twice the number of overlapping lesion detection accuracy, sensitivity was 0.93
accuracy of CS versus DLS was evaluated by pixels between DLS and CS divided by the (excellent), specificity was 0.88 (excellent),
the following 2 methods: sum of the numbers of pixels in DLS and CS. positive predictive value was 0.87 (excellent),
1. Lesion detection accuracy using a The more overlapping between the algorithm’s and negative predictive value was 0.93
dichotomized outcome of “presence of lesion” segmentation and the manual segmentation, (excellent), respectively.
and “absence of lesion” per lesion per root of the higher the DICE.
CS compared with DLS: CS was determined
by the 3 independent observers and based on Voxel-matching Accuracy (DICE)
RESULTS
the probability index by Reit and Gro €ndahl28. The mean DICE index of lesion was used to
From a segmented image by DLS, the The data set contained CBCT images from 6 optimize our DL algorithm, which is 0.52 over
presence of lesion or absence of lesion at each male and 14 female patients ranging from 19– the 5-fold cross validation (standard deviation
root were naturally known and were put into a 71 years old with a mean age of 44.4 years. All 5 0.15, range 5 0.36–0.70). The 2 false
confusion matrix in comparison with a CS reported results were based on cross negatives were the main source of variability for
image. Based on the confusion matrix, validation. DLS was able to distinguish the range of DICE because their respective
sensitivity, specificity, positive predictive value, between the 5 labels described previously DICEs are 0. The mean values of the DICE
and negative predictive value were calculated (Fig. 3A–D). indexes for the other labels were tooth
and graded according to the ranking for structure 5 0.74, bone 5 0.78, restorative
diagnostic tests by Leonardi Dutra et al7 with Lesion Detection Accuracy materials 5 0.58, and background 5 0.95,
scores .80% considered excellent outcomes, CS identified a total of 61 roots within all 20 respectively. The cumulative DICE index for all
between 70% and 80% good, between 60% CBCT volumes. On 29 of 61 roots (47.5%), the actual true lesions was 0.67. Table 1 details all

FIGURE 2 – DLS. Architecture of the DL algorithm for CBCT segmentation and lesion detection. Five slices from each CBCT volume were included in training, with individual U-Nets
built on each slice with shared parameters. These individual U-Nets were integrated by a 3D convolution layer in the end to segment the central slice. The DICE index of “lesion” on the
central slice averaged over 5-fold cross validation was used to optimize the parameters of this DLS design.

JOE  Volume -, Number -, - 2020 Artificial Intelligence for Detecting Periapical Lesions 3
FIGURE 3 – A case example of CS versus DLS demonstrated using a center slice. (A ) RAW: original CBCT slice. (B ) CS: rendering of the clinician-based segmentation. (C ) DLS:
intermediate step of multilabel segmentation applying different weights to individual labels. (D ) DLS: final multilabel segmentation after applying weights, anatomic constraints, and
additional data from adjacent CBCT slices. The red label represents “lesion.” Note the comparison with the RAW image and CS. DICE index for this lesion 5 0.88.

DICE indexes for actual true lesions and still uncertain. The application of an automated favor microbial disinfection before it
individual labels for all samples. detection algorithm for CBCT images is progresses to advanced stages. Many CBCT
significant for a number of reasons. scans are taken to evaluate a single tooth.
Early detection of a PL may allow for However, adjacent teeth contained in the
DISCUSSION timely testing and a clinical differential CBCT volume may demonstrate PLs that
This study investigated the use of artificial diagnosis with a potential of initiating might go unnoticed, particularly in situations in
intelligence, in particular DL, for the automated appropriate treatment. The majority of PLs which clinical evaluations such as thermal or
detection of PLs in a localized CBCT represent apical periodontitis, and early electric pulp testing deliver questionable
environment. Most PLs are indicative of apical detection may increase treatment success. results (eg, the presence of porcelain-fused
periodontitis and thus do not represent a life- Apical periodontitis not only is a predictor of restorations).
threatening disease unless they develop into a success and failure20, but also it was shown Artificial intelligence can reduce
severe acute apical abscess. However, apical that teeth associated with larger apical interpretation bias. DL strategies
periodontitis is very prevalent, and the periodontitis show higher amounts of bacterial demonstrated superior performance over
implications on a patient’s general health are endotoxins29. Thus, earlier treatment may conventional algorithms in biomedical image

TABLE 1 - An Overview of the Roots and Lesions Contained in the 20 Training and Validation Samples as well as DICE Indexes Overall per Individual True Lesion and Overall for Each
Label as Detected by Cone-beam Computed Tomographic Segmentation and Labeling Using the Deep Learning Approach

DICE index per actual true


lesions DICE index overall per label
Roots Roots Roots
per with without Lesion Lesion Lesion Tooth Restorative
Sample sample lesion lesion 1 2 3 Lesion structure Bone materials Background
#1 3 1 2 0.93 0.57 0.83 0.8 0.39 0.96
#2 4 1 3 0.93 0.57 0.79 0.85 0.68 0.94
#3 2 1 1 0.74 0.7 0.66 0.73 0.31 0.95
#4 2 2 0 0.76 0.78 0.69 0.72 0.86 0.55 0.95
#5 3 1 2 0.91 0.76 0.78 0.88 0.77 0.95
#6 3 2 1 0.87 0.78 0.67 0.42 0.48 0.56 0.94
#7 3 1 2 0.71 0.15 0.73 0.84 0.61 0.94
#8 4 2 2 0.94 0.25 0.76 0.7 0.87 0.66 0.93
#9 4 2 2 0.92 0.84 0.68 0.76 0.84 0.18 0.95
#10 4 3 1 0.84 0.86 0 0.51 0.78 0.82 0.76 0.95
#11 3 1 2 0.85 0.52 0.61 0.79 0.61 0.93
#12 2 1 1 0.74 0.15 0.42 0.75 0.6 0.96
#13 4 1 3 0.91 0.49 0.83 0.68 0.68 0.94
#14 2 1 1 0 0 0.91 0.63 0.72 0.94
#15 4 2 2 0.45 0.65 0.43 0.87 0.82 0.61 0.95
#16 3 2 1 0.3 0.42 0.31 0.76 0.8 0.62 0.95
#17 3 2 1 0.72* 0.64 0.76 0.64 0.72 0.97
#18 2 1 1 0.11 0.07 0.6 0.77 0.34 0.93
#19 2 1 1 0.19 0.14 0.79 0.79 0 0.97
#20 4 1 3 0.6 0.54 0.73 0.61 0.09 0.92
Total roots 61 29 32
Cumulative 0.67 0.52 0.74 0.78 0.58 0.95
DICE
index

*One large lesion surrounding 2 roots.

4 Setzer et al. JOE  Volume -, Number -, - 2020


analysis30,31. DL can automatically learn detection accuracy, even if the samples were detection accuracy, the DICE index ratios for
rich features from an image. These range not selected for pristine image quality. Only voxel-matching accuracy, particularly for
from those perceivable by people, such as CBCT volumes with severe beam hardening segmentation of the labels lesion and
intensity, shape, and texture, to those deeply artifacts (eg, from dental implants) were restorative materials need improvement. The
embedded in the image and beyond human excluded. Figure 3A exhibits an example of the reasons for a relatively lower DICE for these 2
perception30,31. Thus, DL holds great promise DLS capability to analyze images of moderate labels were that the sizes of lesions were much
to overcome the CBCT image complexities quality. smaller than the other labels and that a variety
that complicate interpretation. In this limited Besides the detection of PLs, this DL of different materials were combined into a
CBCT environment, our DL algorithm achieved platform may have the potential to detect other single label that inherently differed in shape,
excellent lesion detection accuracy with a oral lesions (eg, for automated screening of location, brightness, and texture. These
sensitivity of 0.93 and a specificity of 0.88, CBCT images for early detection of limitations will be addressed by increasing the
which was higher than achieved by a nonodontogenic lesions, which, if overlooked training size, incorporating anatomic domain
convolutional neural network approach by human observers, may result in significant knowledge into the DL design, and so on. The
evaluating panoramic radiographs for PLs treatment consequences at later stages of the multilabel approach does aid in this goal by
(sensitivity 5 0.65, specificity 5 0.87)14. Our disease or medicolegal implications for the facilitating the baseline architecture of a true 3D
results were comparable with the cumulative practitioner). The long-term goal is the U-Net algorithm26. In addition, overall topology
lesion detection accuracy by clinicians using development of the advanced platform AIDen estimation, location probability analysis, and a
CBCT imaging (sensitivity 5 0.95, specificity 5 seamlessly integrated with clinical workflow domain knowledge approach extended to a
0.88) reported by a systematic review7. and providing decision support for clinicians. 3D setting in combination with transfer learning
However, the latter report included several U-Net architectures are fast27 and would thus from pretrained DL on ImageNet32 and data
in vitro studies under idealized conditions. allow for time-efficient clinical integration. augmentation by image rotation and intensity
DLS produced 2 false-negative reports. Limitations of this study are the transformation can be incorporated. These
One lesion missed was only minimally larger comparison to CS, allowing for human error9. modifications, in combination with a larger
than twice the width of the periodontal Histology or high-resolution micro–computed sample size derived from incremental samples
ligament; the other was indeed detected by tomographic images could be used for added into the training data, can be used to
DLS at first but eliminated in the final additional validation. Also, a direct comparison further improve both lesion detection and
postprocessing steps. This showed that DLS to studies reporting the area under the receiver voxel-matching accuracy.
achieved high consistency compared with CS operating characteristic curve (AUC) may be In the long-term, the addition of
and superiority versus single slice-based limited. AUC calculation relies on the prediction automated image analysis based on 3D
segmentation, even if a full-fledged 3D probability of individual labels derived from imaging may aid clinicians with lesion detection
algorithm, including all volume slices, was not probability vectors originating from the raw and, along with previous investigations,
yet used. Although conventional DL strategies output of DL algorithms. The advanced differential diagnosis of periapical
require several thousand annotated training algorithm used in this study applies cleaning pathology17,33–35 and/or nonodontogenic
samples, a U-Net architecture uses a network processes based on anatomic constraints, lesions.
and training strategy that applies data making an AUC calculation unsuitable because
augmentation to use available annotated no raw output probability vectors remain after
samples at high efficiency27. Thus, even a label voxel fixation. Hence, the DICE index was
ACKNOWLEDGMENTS
small sample size of 20 image sets containing used in this study. Although DLS achieved The authors deny any conflicts of interest
61 roots could produce excellent lesion excellent sensitivity and specificity for lesion related to this study.

REFERENCES
1. Becconsall-Ryan K, Tong D, Love RM. Radiolucent inflammatory jaw lesions: a twenty-year
analysis. Int Endod J 2010;43:859–65.

2. €
Eriksen HM. Epidemiology of apical periodontitis. In: Orstavik D, Pitt Ford TR, editors. Essential
Endodontology. Oxford, UK: Blackwell Science; 1998. p. 179–91.

3. Bender IB, Seltzer S. Roentgenographic and direct observation of experimental lesions in bone: I.
1961. J Endod 2003;29:702–6.
4. Goldman M, Pearson AH, Darzenta N. Endodontic success–who’s reading the radiograph? Oral
Surg Oral Med Oral Pathol 1972;3:432–7.

5. Goldman M, Pearson AH, Darzenta N. Reliability of radiographic interpretations. Oral Surg Oral
Med Oral Pathol 1974;38:287–93.

6. Setzer FC, Hinckley N, Kohli MR, et al. A survey of CBCT use amongst endodontic practitioners in
the United States. J Endod 2017;43:699–704.
7. Leonardi Dutra K, Haas L, Porporatti AL, et al. Diagnostic accuracy of cone-beam computed
tomography and conventional radiography on apical periodontitis: a systematic review and meta-
analysis. J Endod 2016;42:356–64.

JOE  Volume -, Number -, - 2020 Artificial Intelligence for Detecting Periapical Lesions 5
8. Kruse C, Spin-Neto R, Evar Kraft DC, et al. Diagnostic accuracy of cone beam computed
tomography used for assessment of apical periodontitis: an ex vivo histopathological study on
human cadavers. Int Endod J 2019;52:439–50.

9. Parker JM, Mol A, Rivera EM, et al. Cone-beam computed tomography uses in clinical
endodontics: observer variability in detecting periapical lesions. J Endod 2017;43:184–7.

10. Suzuki K. A review of computer-aided diagnosis in thoracic and colonic imaging. Quant Imaging
Med Surg 2012;2:163–76.
11. Firmino M, Angelo G, Morais H, et al. Computer-aided detection (CADe) and diagnosis (CADx)
system for lung cancer with likelihood of malignancy. Biomed Eng Online 2016;6:2.

12. Aamodt A, Kvistad KA, Andersen E, et al. Determination of Hounsfield value for CT-based design
of custom femoral stems. J Bone Joint Surg Br 1999;81:143–7.

13. Prevrhal S, Engelke K, Kalender WA. Accuracy limits for the determination of cortical width and
density: the influence of object size and CT imaging parameters. Phys Med Biol 1999;44:751–64.
14. Ekert T, Krois J, Meinhold L, et al. Deep learning for the radiographic detection of apical lesions. J
Endod 2019;45:917–22.

15. Abdolali F, Zoroofi RA, Otake Y, et al. Automatic segmentation of maxillofacial cysts in cone beam
CT images. Comput Biol Med 2016;72:108–19.

16. Li S, Fevens T, Krzyzak A, et al. Toward automatic computer aided dental X-ray analysis using
level set method. Med Image Comput Comput Assist Interv 2005;8:670–8.
17. Okada K, Rysavy S, Flores A, et al. Noninvasive differential diagnosis of dental periapical lesions
in cone-beam CT scans. Med Phys 2015;42:1653–65.

18. Hiew L, Ong S, Foong KW. Tooth segmentation from cone-beam CT using graph cut. In:
Proceedings of the Second APSIPA Annual Summit and Conference. Singapore: ASC; 2010.
p. 272–5.

19. Berdouses ED, Koutsouri GD, Tripoliti EE, et al. A computer-aided automated methodology for
the detection and classification of occlusal caries from photographic color images. Comput Biol
Med 2015;62:119–35.

20. Strindberg LZ. The dependence of the results of pulp therapy on certain factors; an analytic study
based on radiographic and clinical follow-up examinations. Acta Odontol Scand 1956;14(Suppl
21):1–175.

21. Abella F, Patel S, Duran-Sindreu F, et al. Evaluating the periapical status of teeth with irreversible
pulpitis by using cone-beam computed tomography scanning and peri-apical radiographs. J
Endod 2012;38:1588–91.

22. Zhang MM, Liang YH, Gao XJ, et al. Management of apical periodontitis: healing of post-
treatment periapical lesions present 1 year after endodontic treatment. J Endod 2015;41:1020–
5.

23. Schloss T, Sonntag D, Kohli MR, et al. A comparison of two- and three-dimensional healing
assessment after endodontic surgery using CBCT volumes or periapical radiographs. J Endod
2017;43:1072–9.

24. Ma H, Poly A, Sehgal A, et al. Volumetric assessment of sinus membrane dimensions in relation
to endodontically treated and healthy teeth. J Endod 2018;44:e21.
25. Poly A, Marques F, Sassone L, et al. The ability of four different instrumentation systems in
shaping oval canals: a micro-computed tomographic analysis. J Endod 2018;44:e40.
26. € Abdulkadir A, Lienkamp SS, et al. 3D U-Net: learning dense volumetric segmentation
Çiçek O,
from sparse annotation. Med Image Comput Comput Assist Interv 2016;9901:424–32.

27. Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image
segmentation. arXiv 2015;1505:04597.
28. €ndahl HG. Application of statistical decision theory to radiographic diagnosis of
Reit C, Gro
endodontically treated teeth. Scand J Dent Res 1983;91:213–8.
29. Gomes BP, Endo MS, Martinho FC. Comparison of endotoxin levels found in primary and
secondary endodontic infections. J Endod 2012;38:1082–6.

30. Hosny A, Parmar C, Quackenbush J, et al. Artificial intelligence in radiology. Nat Rev Cancer
2018;18:500–10.
31. Shen D, Wu G, Suk HI. Deep learning in medical image analysis. Annu Rev Biomed Eng
2017;19:221–48.

6 Setzer et al. JOE  Volume -, Number -, - 2020


32. Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural
networks. Adv Neural Inf Process Syst 2012:1097–105.

33. Simon JH, Enciso R, Malfaz JM, et al. Differential diagnosis of large periapical lesions using cone-
beam computed tomography measurements and biopsy. J Endod 2006;32:833–7.
34. Guo J, Simon JH, Sedghizadeh P, et al. Evaluation of the reliability and accuracy of using cone-
beam computed tomography for diagnosing periapical cysts from granulomas. J Endod
2013;39:1485–90.
35. Chanani A, Adhikari HD. Reliability of cone beam computed tomography as a biopsy-
independent tool in differential diagnosis of periapical cysts and granulomas: an in vivo study. J
Conserv Dent 2017;20:326–31.

JOE  Volume -, Number -, - 2020 Artificial Intelligence for Detecting Periapical Lesions 7

You might also like