CBCT has become very important radiographic technique in dentistry. Use of CBCT in dental procedures has gained popularity in recent years due to its low cost, fast image production rate, and lower radiation dose in comparison to medical CT [39]. However, CBCT machines are operated at milliamperes that roughly one order of magnitude below the medical CT machines. Noise is defined as an unwanted disturbance of a signal that tends to obscure the signal's information. Despite the reduction in the radiation dose, a high noise level or lower signal-to-noise ratio is expected in CBCT images. Noise reduces contrast resolution and affects the ability to segment low-density tissues effectively [40, 41]. Artifact is any distortion or error in the image that is unrelated to the subject. Image artifacts are one of the drawbacks of the clinical use of CBCT. Artifacts may obscure or simulate pathology of the head and neck region, including dental caries [39, 41].
Scatter is caused by those photons that are diffracted from their original path after interaction with matter. The scattered photons are captured by the sensor and simply added to the primary intensity. Geometry of the detector is an important factor for this image-degrading effect of scattered radiation, as the sensor gets larger, the probability of catching a scattered photon is raised. Scatters reduce soft-tissue contrast and affect the density of all tissues [40]. The streak artifacts caused by scatter are very similar to those of beam hardening [40, 41]. Beam hardening is one of the most common sources of artifacts. As the beam passes through the object, a highly absorbing material in the object, such as metal, can function as a filter to absorb the lower energetic photons more rapidly than the higher-energetic photons. Hence, the beam spectrum becomes rich in high-energy photons and the mean energy increases. When the spectrum of the captured beam contains relatively more higher-energetic photons than the emitted ray, the beam becomes ‘hardened’ and an artifact is induced, resulting in dark streaks [39, 40]. Artifacts are related to several factors such as the object, material type, FOV, imaging device and parameters [42, 43]. The effectiveness of metal artefact reduction algorithm is investigated by several authors [44–46]. Xie et al. proposed a deep convolutional neural network to reduce scatter artifacts for CBCT in an image-guided radiation therapy system [47].
Several authors investigated the potential of using CBCT instead of plain radiographs in detecting dental caries. Studies on this issue have reported varying results, perhaps due to differences in methodology. Young et al. evaluated the CBCT images in detecting proximal and occlusal caries by mounting 146 non-restored extracted human teeth in plaster. Caries lesions are categorized according to location and depth, and practicing dentists are found to be more successful in CBCT images with the average sensitivity of 0.61 when compared to plain radiographs, but not occlusal caries [48]. Kayipmaz et al. investigated the use of CBCT in detecting occlusal and approximal caries using 72 extracted human teeth. In their study, CBCT is reported to be superior in detecting not the approximal, but the occlusal caries, when compared to plain radiographs [49]. Krzyżostaniak et al. conducted a study using 135 extracted human posterior teeth and accuracy of detecting non-cavitated proximal caries with CBCT unit is reported to be (0.629) inferior than other intraoral radiography techniques. However, CBCT system is reported to be slightly better for detecting occlusal carious lesions [50].
Unlike researchers that included occlusal caries, some studies excluded this location, as we have, and focused on approximal caries detection. Zhang ZL et al. evaluated 39 non-cavitated and unrestored human permanent teeth for approximal caries. The mean ROC values for two different CBCT devices are reported to be 0.528 and 0.525 (p = 0.763). The performance of CBCT is reported to be a little better than chance when compared to plain radiography [51]. Valizadeh et al. embed 84 extracted human teeth in blocks and the area under the ROC curve, sensitivity, specificity, accuracy, positive and negative predictive values of CBCT images are reported to be 0.568, 0.835, 0.637, 0.714, 0.598 and 0.856, respectively. Afterall, CBCT images did not enhance the detection of proximal caries in comparison with plain radiography [52]. Wenzel et al. mounted 257 non-filled human teeth in plaster to be evaluated and found that CBCT was more accurate than intraoral radiography in detecting approximal caries [53].
Several studies are performed with the motivation that the artifacts caused by restorative materials may affect the diagnosis of caries. Charuakkra et al. compared CBCT and bitewing radiographs in detecting secondary caries using 120 cavity slots with different restorative materials. The mean ROC values for the CBCT system are reported to be 0.995, and 0.978, making CBCT superior to bitewing radiographs [54]. Melo et al. evaluated the use of CBCT in detecting recurrent caries-like lesions created artificial under restorative materials. In their study, CBCT and intraoral radiography are found to be similar in detecting demineralization under restorations [55]. In addition, not all CBCT machines are duplicates due to the adoption of different production technologies. Considering that differences in production technology may affect the diagnosis of dental caries, Qu et al. investigated the effect of two different detector types employed in five CBCT systems in diagnostic accuracy of approximal carious lesions by evaluating 78 approximal surfaces. According to the results of this study, the differences between five different CBCT devices and two different detector types were not found to be statistically significant [56].
In this study, the areas under ROC curves are found to be better than Zhang ZL et al. and Valizadeh et al. and close to the Charuakkra et al. [51, 52, 54]. Sensitivities are found better than the research of Young et al. and similar to the study of Valizadeh et al. [48, 52]. The overall consensus is increased in aided evaluations when compared to unaided evaluations. It can be suggested that the software aid might have an impact on the magnitude of the changes in the observer results. It was found that the third observer made the most changes between aided and unaided evaluations, resulting in the highest increase in accuracy.
Cardoso et al. reported that gold standard data or method is related to something that has already been checked (histologically, microscopically, chemically, etc.) and presents the best accuracy (sensitivity and specificity). Ground truth means data and/or methods related to more consensus or reliable values/aspects that can be used as references, but were not or cannot be checked [57]. In this retrospective study, the golden standard examinations could not be performed, however, the ground truth was determined by reaching to a consensus among observers under supervision of an experienced dentomaxillofacial radiologist. The ground truth in this study should not be considered as ‘the golden standard’, but as more consensus and reliable values/aspects than the observers’ single evaluations. It may be useful to reiterate that this study is about radiographic caries detection, not caries diagnosis. Another limitation of the study is that the status of restorative materials and artifacts in the images were not taken into consideration. An accurate analysis would not be possible in this regard, as there are many parameters (material size, number, relative positions, etc.) that affect artifacts in CBCT images. Evaluating the effect of CBCT artefacts on performance of the machine learning models can be the subject of further studies. Lastly, CBCT imaging require relatively higher radiation dose, and its use for caries detection is not justified. However, once the CBCT volume is acquired for other reasons, it already contains information for detecting the tooth caries signs. We suggest that it may be beneficial not to miss such information with the machine learning approach in the evaluation of these images.
In this study, radiographic evaluations performed by three observers were found to be more compatible and accurate with the aid of the AI system when compared to the evaluations without, in detecting dental caries on CBCT images. Our study does not recommend justifying the use of CBCT imaging for caries diagnosis, but suggests that once the volumetric data is acquired, machine learning tools can be helpful in detecting the caries signs. As technology advances, integration of similar tools into the digital radiology workflow can assist clinicians in evaluating radiographic data.