Evaluation of children’s oral diagnosis and treatment using imaging examination using AI based Internet of Things

2.1Study subjects

82 children with CDF who were diagnosed and treated at the Yantai Mountain Hospital from December 2020 to March 2022 were selected as the study subjects. All children underwent CBCT, and the degree of dental anxiety was assessed using the children’s fear survey schedule-dental subscale (CFSS-DS) and Venham scale. Inclusion criteria: (1) children aged 3–6 years; (2) children with good general health; (3) children need to return to the dental clinic for many times; (4) CFSS-DS score higher than 34 points, (5) Venham grade ⩾ 3; (6) American Society of Anesthesiologists (ASA) physical assessment grade I–II, children without GA contraindications; (7) children without no mental disorders or cognitive impairment; exclusion criteria: (1) ASA grade ⩾ III, GA contraindications; (2) oral diseases complicated by acute inflammation; (3) children with cardiovascular system diseases; (4) children with respiratory diseases; (5) children with mental or mental disorders; (6) children taking anxiolytic and sedative drugs within 7 days before treatment; (7) mild anxiety children whose treatment was not affected; (8) parents refused to participate in this trial; the family members of the children signed informed consent, and the experimental process was approved by the ethics committee of Yantai Mountain Hospital.

2.2Dental biomedical signal and image preprocessing

CBCT images during oral diagnosis and treatment in Yantai Mountain Hospital were collected and saved as DICOM standard format. All CBCT images were scanned by Italian NEWTOM VGi scanner. The scanning parameters were as follows: tube voltage 110 kV, current 2.81 mA, scanning field of vision 150 mm × 150 mm, scanning time 6,000 ms, resolution 0.1 mm × 0.1 mm, axial layer thickness 0.3 mm.

For DICOM biological signal data, image gray value transformation was required. The specific operation method was as follows: window-adding method [19] was adopted to display the gray scale of DICOM format image, and linear mapping between the original gray scale value and the displayed gray scale was established according to the gray scale value of image. It establishes a linear mapping between the original grayscale values and the displayed grayscale. By selecting a specific range of grayscale values, called a window, and mapping it to the full display range, the method enhances the visibility of relevant image details. This allows healthcare professionals to better visualize anatomical structures or pathological features. The implementation of the window-adding method may vary depending on the software or viewer used, and different users may customize the window settings based on their specific requirements. Then, the full gray value of the original image can be expressed as follows.

O is the original gray value of the image, V is the displayed gray value of the image, W is the window width, and C is the window level.

The logarithmic transformation method of window function was adopted to map the gray scale of the image, and the gray scale value displayed in the window CBCT image is computed.

The CBCT image after gray value transformation needed to be processed by adaptive enhancement algorithm to improve the gray level balance of the image. CLAHE enhancement algorithm was used to process the image. CLAHE is an image enhancement technique that improves contrast and quality of digital images by dividing them into smaller tiles and applying histogram equalization to each tile individually. Unlike traditional histogram equalization, CLAHE applies the equalization to a limited range of intensity values within each tile and uses a clipping or limiting function to prevent over-enhancement and noise amplification. By combining the enhanced tiles, CLAHE produces an image with improved contrast and quality while avoiding artifacts and noise. The gray value f⁢(x,y) of the processed image is presented as below.

W is a two-dimensional image template, and m and n are stretching factors.

2.3Establishment of CBCT image intelligent segmentation model based on level set algorithm

CV model can accurately segment image edges and obtain global convergence, but it still has obvious defects when segmenting images with uneven gray distribution. However, they have significant limitations for images with an uneven gray distribution. Such images present difficulties for CV models, resulting in errors such as inaccurate edge detection, boundary placement, and sensitivity to local variations. The segmentation performance of these models is being actively enhanced by researchers in order to better handle uneven gray distribution and improve segmentation accuracy [20]. For the CBCT image after adaptive enhancement processing, if CV is used for processing, its level set energy function E⁢(θ) is computed.

Eg is the global driving energy function, El is the local driving energy function, γ is the global driving energy function parameter, and τ is the local driving energy function parameter.

Based on global information, Eg can increase the image contour and robustness of noise.

B1 is the average gray level in the evolution curve, B2 is the average gray level outside the evolution curve, Hβ⁢[θ⁢(x)] is a regularized Heaviside function.

Local binary fitting (LBF) model can solve the problem of image segmentation accuracy with uneven gray scale, but it is highly dependent on the initial contour position [21]. LBF model is introduced into the local driving energy function of CV model to avoid the shortcomings of CV model in segmenting images with uneven gray distribution. The accuracy of the Local Binary Fitting (LBF) model for image segmentation in the presence of uneven grayscale depends greatly on the initial contour position. The model iteratively refines the contour based on local image information and prior knowledge of the object shape. Howeve an incorrect initial contour placement can lead to inaccurate segmentation. Therefore, selecting the optimal initial contour position is crucial to achieve precise segmentation results.

The regularization term is introduced in Eq. (6) to constrain the shape change of root canal segmentation contour, so the level set energy functions are expressed as follows.

L⁢(θ) is the length penalty item; P⁢(θ) is the energy penalty term, Er⁢(θ) is a regularized item, μ is control parameter for the length penalty item, σ is control parameters for energy penalty item.

The specific process of CV-LBF model segmentation of tooth CBCT image is shown in Fig. 1.

Figure 1.

Flowchart of CBCT images of teeth segmented by CV-LBF model.

First, the windowing method is applied to transform the image gray value, the median filter is used to denoise the image, and the CLAHE enhancement algorithm to improve the gray balance of CBCT image. The median filter reduces noise in images by replacing each pixel with the median value of its neighboring pixels, making it effective at removing isolated noisy pixels without affecting the image details and edges. Its adaptability to the local image structure, rather than a fixed kernel size, helps preserve the image quality. An appropriate initial layer image is selected to obtain a small window area containing the target teeth. The level set segmentation method is adopted to perform the initial layer results. In children’s oral diagnosis and treatment, level set segmentation is utilized to segment the oral cavity and teeth from other structures in the image. This involves image acquisition, preprocessing, initialization, evolution, and postprocessing to extract a final contour for diagnosis and treatment planning. The resulting segmented image can be analyzed to detect abnormalities and plan orthodontic treatments or oral surgeries. Based on the initial layer segmentation, the CV-LBF tooth segmentation model in the direction of the crown and root respectively for fine segmentation, the target tooth segmentation image is obtained.

2.4CBCT image evaluation index

The performance of CBCT images was evaluated with segmentation accuracy (SA), maximum symmetrical surface distance (MSSD), average symmetrical surface distance (ASSD), over segmentation rate (OR), and under segmentation rate (UR). Maximum Symmetrical Surface Distance (MSSD) and Average Symmetric Surface Distance (ASSD) are two metrics that assess the precision of image registration algorithms by measuring the distance between two surfaces in a symmetric manner. They share the feature of being used in medical imaging to align images of the same patient taken at different times or align images from different patients to a common coordinate system. However, the main contrast between them is that while MSSD computes the maximum distance between the two surfaces, ASSD measures the average distance. The segmentation result of optimized CV-LBF algorithm was compared with those of adaptive threshold [22], region growing [23], DRLSE [24] and CV algorithms [25]. Adaptive thresholding is an image processing technique that adjusts the threshold for each pixel or region based on local image properties, useful for enhancing visibility and segmenting regions of interest in CBCT images. Region growing is a segmentation algorithm that groups neighboring pixels or voxels based on similarity, allowing for the extraction of distinct regions in CBCT images. Distance Regularized Level Set Evolution (DRLSE) is a level set-based segmentation algorithm that accurately captures complex shapes and boundaries of anatomical structures in CBCT images and finally computer vision algorithms encompass various techniques used for tasks such as image registration, feature extraction, object detection, and classification in CBCT image evaluation, enabling tasks like image alignment, quantitative analysis, and automated quality control.

R is the tooth or root canal area of the physician manually segmented image; T is the area of the intelligent segmented image, S is the area of the intelligent segmented image extended outside the R contour, and U is the area of the intelligent segmented image contracted into the R contour. When O⁢R and R⁢U< 10% or S⁢A> 80%, it is considered that the tooth image segmentation is qualified, and the smaller the MSSD and ASSD values, the closer the computer results are to the manual results, and the better the segmentation results are.

2.5Intervention modalities in children with CDF

According to the voluntary principle, the family members of cldren with CDF chose the intervention method, of which 38 children were in the general anesthesia group (GAG) and the remaining 44 children chose the protective fixation intervention method and they were in controls. Before treatment, the two groups of children through the cartoons, toys, doctors dress, and other means to establish a warm and relaxed treatment environment, preoperative nurses through chatting with children and other means to promote a sense of trust, making the children’s mood relaxed.

The child in controls was fasted and water deprived for more than 4 hours before surgery, and the family signed the informed consent form for protective fixation therapy. The nurse assisted the child to lie supine on the oral treatment chair and adjust it to the appropriate height, and the restraint plate was placed flat on the oral treatment chair to fix the top of the restraint plate and the headrest. The child was wrapped with a restraint plate towel to expose only the child’s head and neck. During oral treatment, the family assisted with head immobilization and mouthpiece was used to assist the child’s mouth opening.

Children in GAG were fasted and water deprived for 8 hours before anesthesia. All children underwent GA using intravenous inhalation combined anesthesia technique. Combining inhalation anesthetics is common in anesthesia practice because it allows for a more precise and reliable level of anesthesia while reducing the risk of side effects associated with higher concentrations of a single agent. This approach also enables synergistic effects and reduced side effects, as well as flexibility in tailoring the anesthetic to the individual patient and procedure. After entering the operating room, the child underwent sedation, arteriovenous puncture, GA, and other operations. The assistant recorded the intraoperative conditions in detail. At the end of the operation, the patient was transferred to the GA recovery room and could leave the hospital after the vital signs were stable for 2 hours.

2.6Outcome measures

CFSS-DS and Venham [26] scores were counted before intervention in both groups, and CFSS-DS scores were also recorded at 2 hours after intervention in controls, after recovery in GAG. CFSS-DS and Venham scores were performed in all children at return visit 1 week after surgery.

The CFSS-DS [27] was rated 1 to 5 points according to the degree of fear, 5 points for very fear, 4 for more fear, 3 for comparative fear, 2 for a little fear, and 1 for no fear, with a total of 17 items, and with a total score of more than 42 indicating a high level of fear and less than 42 indicating a low level of fear.

2.7Statistical methods

Data were statistically analyzed using SPSS 22.0. The included measurement data were in accordance with the normal distribution, the measurement data were presented as mean ± standard deviation (x¯±s), and the variables between groups were compared adopting the independent sample t-test; percentage (%) indicated the enumeration data, and the chi-square test was applied. Statistical significance was defined as P< 0.05.

3.1CBCT image segmentation results for CV-LBF model teeth

The CV-LBF model segments root canals in CBCT images by extracting local binary features and training a machine learning algorithm to predict root canal boundaries. It is trained on annotated datasets and can be used to segment the root canals of individual or multiple teeth, as well as images taken at different angles. CV-LBF model was applied to segment the root canals of teeth at different positions in CBCT images, and the results are illustrated in Fig. 2. CV-LBF model can accurately segment the root canals of teeth with different morphologies.

Figure 2.

Automatic segmentation diagram of root canal of CV-LBF model.

The CV-LBF model (red) was compared with the manual segmentation (green) (Fig. 3), and the tooth segmentation contours of the CV-LBF model were highly coincident with the manual segmentation results.

Figure 3.

CBCT cross-sectional tooth segmentation results. A: original CBCT image; B: molar segmentation result; C: incisor segmentation result.

3.2Comparison of tooth segmentation accuracy of intelligent algorithm

The performance of various image segmentation algorithms on CBCT images, measured using the S⁢A (segmentation accuracy) value. The algorithms evaluated were CV-LBF, adaptive threshold, region growing, DRLSE, and CV. The results show that the CV-LBF algorithm achieved the highest S⁢A value of (89.92 ± 8.25) %, followed by DRLSE (68.20 ± 6.94) %, region growing (78.63 ± 7.71) %, CV (81.52 ± 6.98) %, and adaptive threshold algorithms (74.86 ± 6.77) %. Among these, the difference in the S⁢A values of the CV-LBF and region growing algorithms was statistically significant (P< 0.05), indicating that the former performed better than the latter (Fig. 4).

Figure 4.

Comparison of S⁢A values of CBCT images of teeth segmented by different algorithms. ^* indicates statistical difference compared with region growing algorithm, P< 0.05.

The OR value is a common metric used to evaluate the similarity between the segmented object and the ground truth object. It represents the ratio of the intersection between the segmented object and the ground truth object to the union of the two objects. A higher OR value indicates a better segmentation accuracy. The results find that the OR values of tooth CBCT images segmented by the CV-LBF (1.53 ± 0.11) %, region growing (5.56 ± 0.34) %, adaptive threshold (3.27 ± 0.26) %, DRLSE (1.84 ± 0.15) %, and CV (4.16 ± 0.34) % algorithms were lower compared to the OR values of tooth CBCT images segmented by the region growing and CV algorithms. Specifically, the OR value of the CV-LBF algorithm was significantly lower than the OR value of the region growing algorithm (P< 0.05), which indicates that the CV-LBF algorithm outperformed the region growing algorithm in terms of segmentation accuracy (Fig. 5).

Figure 5.

Comparison of OR values of CBCT images of teeth segmented by different algorithms.

U⁢R is a metric used to assess the accuracy of a segmentation algorithm, which calculates the ratio of the union of the ground truth and segmented regions to their intersection. It indicates how much of the true positive region is correctly segmented by the algorithm. CV-LBF (2.36 ± 0.25) %, Region growing (6.61 ± 0.54) %, Adaptive threshold (5.23 ± 0.48) %, DRLSE (3.45 ± 0.29) %, and CV (4.29 ± 0.31) %. The CV-LBF algorithm showed a significantly lower mean U⁢R value compared to the region growing algorithm (p< 0.05), indicating better performance in accurately segmenting dental CBCT images (Fig. 6).

Figure 6.

Comparison of U⁢R values of segmenting CBCT images by different algorithms.

3.3Comparison of MSSD and ASSD values of CBCT image segmentation with intelligent algorithm

The evaluation metric used to compare the algorithms was the Mean Symmetric Surface Distance (MSSD), which measures the average distance between corresponding points on the surfaces of the segmented image and the ground truth image. The CV-LBF, region growing, adaptive threshold, DRLSE, and CV algorithms were compared, in which the MSSD values obtained for each algorithm were (1.98 ± 0.71), (9.35 ± 0.96), (7.39 ± 0.78), (10.26 ± 3.03), and (12.97 ± 2.15) mm, respectively. From the results, it is clear that the CV-LBF algorithm performed better than the CV algorithm, as the MSSD value obtained for CV-LBF was significantly lower (P< 0.05) than that CV (Fig. 7).

ASSD is a widely used metric that measures the distance between the surfaces of two segmented objects. The CV-LBF, regional growth, adaptive threshold, DRLSE, and CV are the five different algorithms tested in the experiment. The average ASSD values obtained for each algorithm were (0.27 ± 0.03), (6.64 ± 0.58), (2.05 ± 0.33), (3.16 ± 0.25), and (5.02 ± 0.48) mm, respectively. The CV-LBF algorithm performed better than the regional growth algorithm, as the ASSD value obtained for CV-LBF was significantly lower (P< 0.05). The CV-LBF algorithm was more accurate in segmenting the CBCT images and produced segmentations that were closer to the ground truth image (Fig. 8).

3.4Comparison of clinical characteristics of children with CDF before intervention

Statistical analysis was performed on the age, gender, parenting style, whether the child was the only child, and mode of delivery of GAG and controls (Table 1). The results indicated that there were no statistically significant differences in the age, gender, parenting style, whether the child was the only child, and mode of delivery between two groups (P> 0.05).

Figure 7.

Comparison of MSSD values of segmenting CBCT images by different algorithms. ^*Indicates statistical difference compared with CV algorithm, P< 0.05.

3.5Comparison of education level of mothers and family income

There were not clear different in maternal education level and family income between the two groups (P> 0.05) (Fig. 8).

Figure 8.

Comparison of family income and maternal education level.

3.6Comparison of distribution of temperament types in children

Figure 9 shows the proportion of different temperament types in the two groups, 9 (23.68%), 10 (26.32%), 11 (28.95%), and 8 (21.05%) children with phlegmatic temperament, biliary temperament, sanguine temperament, and depressive temperament in GAG; 10 (22.73%), 12 (27.27%), 12 (27.27%), and 10 (22.73%) children in controls, respectively (P> 0.05).

Figure 9.

Comparison of temperament type distribution.

Figure 10.

Comparison of CFSS-DS values at different time periods before and after intervention. ^**Means compared with before intervention, P< 0.001; ^##means compared with controls, P< 0.001.

Figure 11.

Comparison of Venham scores before and after intervention.

3.7Comparison of CFSS-DS values at different time periods before and after intervention

Before intervention, the CFSS-DS scores of GAG and controls were (35.47 ± 2.97) and (35.26 ± 3.13) (P> 0.05). The CFSS-DS scores of GAG and controls after 2 hours of intervention were (26.71 ± 2.22) and (38.55 ± 3.46). After 2 hours of intervention, relative to before intervention, the score of GAG was (8.76 ± 5.72) lower, and the score of GAG was (3.29 ± 2.76) higher. After 2 hours of intervention, the score of GAG was lower (P< 0.001), the score of controls was higher (P< 0.001), and the score of GAG after 2 hours of intervention was lower relative to controls (P< 0.001). The score of GAG and controls at return visit after 1 week of intervention were (27.03 ± 2.49) and (39.04 ± 3.37). At return visit after 1 week of intervention, compared with before intervention, the score of GAG was (8.44 ± 6.26) lower, and that of controls was (3.78 ± 3.11) higher; the score of GAG was obviously lower (P< 0.001), and that of controls was higher (P< 0.001); the score of GAG at return visit after 1 week of intervention was lower in contrast with controls (P< 0.001). There was no significant difference in CFSS-DS score at return visits after 1 week of intervention compared with that after 2 hours of intervention between the two groups (P> 0.05) (Fig. 10).

3.8Comparison of Venham scores before and after intervention

There was not obviously different in Venham scores between the two groups before intervention (P> 0.05), Venham scores after intervention were (1.26 ± 0.27) and (3.33 ± 0.29), and after intervention, the score of GAG decreased (P< 0.001), and was lower than that of controls (P< 0.001). In controls, Venham score before the intervention was not clearly different from that after the intervention (P> 0.05) (Fig. 11).

Before the intervention, GAG had 14 (36.84%), 13 (34.21%), and 11 (28.95%) children with Venham grade 3, 4, and 5, 16 (36.36%), and controls had 16 (36.36%), 17 (38.64%), and 11 (25.00%) children (P> 0.05). After intervention, the proportion of children with grade 0, 1, 2 and 3 was higher in GAG than in controls (P< 0.001), while the proportion of children with grade 4 and 5 was higher in controls than in GAG (P< 0.001) (Fig. 12).

Figure 12.

Comparison of Venham grade distribution after intervention. ^**Means P< 0.001 relative to controls.