Deep learning algorithm for the detection and 3D segmentation of maxillofacial bone lesions in CBCT

doi:10.21203/rs.3.rs-1565644/v1

PDF

Research Article

Deep learning algorithm for the detection and 3D segmentation of maxillofacial bone lesions in CBCT

https://doi.org/10.21203/rs.3.rs-1565644/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Objectives: This study aimed to develop an automated deep-learning algorithm for the detection and 3D segmentation of incidental jaw lesions in maxillofacial CBCT scans.

Materials and Methods: The dataset included 82 CBCT scans with and without histologically confirmed benign bone lesions, obtained from three CBCT devices using different imaging protocols. The dataset consisted of axial CBCT images and was divided into training dataset (20,214 axial images), validation dataset (4,530 axial images) and testing dataset (6,795 axial images). A Mask-RCNN based deep-learning algorithm segmented the bone lesion in each axial image. The analysis of sequential slices improved the mask RCNN performance and assisted in classifying each CBCT case as containing bone lesions or not. Thereafter, the algorithm generated a 3D segmentation of the lesions.

Results: The accuracy of the algorithm for classifying each CBCT case as either containing bone lesions or not, was 100%. The algorithm performance for detecting lesions in individual axial images showed high sensitivity (95.9%) and high precision (98.9%).

Conclusions: Our deep learning algorithm can detect incidental bone lesions in CBCT scans with high accuracy, high sensitivity and high precision and is highly recommended for bone lesion detection and follow-up in CBCT imaging.

Clinical relevance: The increasing number of CBCT scans performed worldwide dictates a clinical need for developing an automated tool, which will review, detect and volumetrically demonstrate incidental bone lesions in CBCT scans with high accuracy and low false-positive rate.

CBCT

maxillofacial bone-lesions

Deep learning

Region-Based-Convolutional-Neural-Networks (RCNN)

3D-segmentation

Cone Beam Computerized Tomography (CBCT) is a volumetric imaging modality used in the oral and maxillofacial field for diagnosis and treatment planning of both dental and non-dental pathologies including bone lesions (BL) (1) BL consist of cysts, benign and malignant tumors which may be clinically apparent or may be found incidentally on CBCT. Incidental BL on CBCT are reported in different frequencies and require either immediate intervention or follow-up in up to 45.4% of the cases (2, 3)with 0.3%-1.4% of the cases suspected to be malignant (4, 5). Detection of incidental BL is important to avoid bone expansion (6), dental malocclusion (7) and the development of malignant lesions (8). Moreover early bone lesion detection allows conservative treatment with reduced morbidity and mortality (9).

Artificial Intelligence has been implemented in maxillofacial radiology for detection, classification and segmentation of area of interests or objects in the image for purposes of diagnosis and image analysis (10). Deep learning, a subgroup of machine learning allows the derivation of high-level abstractions and complex features from big data through a combination of several nonlinear transformations using artificial neural networks (ANN) (11). These neural networks are trained to learn from known data by minimizing the error between the network prediction and the data ground-truth during repetitive steps. Convolutional neural network (CNN), a branch of ANN, has been used as a main network architecture in the field of OMF radiology (12).

Many studies have used various AI models and applications in CBCT imaging (13). Different CNN architectures have been used for 3D segmentation (14–16) of various regions of interest in CBCT. Few studies have been published for the classification and segmentation of BL in maxillofacial CBCT. Some researchers focused on the classification of the lesion type while requiring manual or a semi-automatic segmentation of the lesion contour (17, 18). Another study, required manual marking of a rectangle containing the lesion in a specific slice for the classification of BL by CNN (19). Some studies, focused on automatic 3D segmentation of only periapical lesions (20–22) while other studies used asymmetry analysis BL detection (23) and Support Vector Machine (SVM) as well as Sparse Discriminant Analysis (SDA) with spherical harmonics features for classifying the lesion type (24) .

BL may be found in both jaws, appear in different sizes and shapes and demonstrate various abnormal densities, which radiographically may be either radiolucent, radiopaque or mixed. Moreover, the scan parameters for CBCT examinations, using various CBCT machines and different protocols may vary in their voxel size, field of view and the anatomical location of jaws being scanned (25). None of the available studies provided fully-automated detection of BL, irrespective of their location in the jaws and CBCT scan parameters.

In this study, we developed a prototype of a fully-automated generic deep learning algorithm designed for the automated detection and 3D segmentation of various BL in CBCT scans irrespective of scan parameters and machines.

A. THE DATASET

This retrospective study included 41 CBCT scans with BL and 41 without bone lesions (w/o BL) performed between January 2019 and December 2020, at our oral maxillofacial imaging unit. This study was approved by the Institutional Review Board (HMO-0297-21). This study was retrospective and involved only medical record analysis, therefore, the IRB approved an exemption from informed consent.

The study group included CBCT scans with well-defined hypodense benign BL, with or without bone expansion, which were performed prior to surgical intervention. All BL were interpretated by a maxillofacial imaging expert (CN) and confirmed by histopathological results determined by microscopic examination of suspected tissue. The control group, w/o BL, consisted of CBCT scans performed prior to dental implant placement, of either the maxilla or the mandible, or both jaws.

CBCT scans with movement artifacts, were excluded from both groups since they are usually non diagnostic. However, scans with metal artefact were not excluded, since metal artefact are very common in CBCT scans and it is important to include them for training the deep learning algorithm. In addition, these artefacts do not usually appear in the same height as the lesions, and when both bone lesions and metal artifacts are visualized in the same slice, the artifact does not significantly modify the appearance of the lesions.

All CBCT scans were assigned into training (50 cases), validation (10 cases) and testing (22 cases) datasets, including 20214, 4530 and 6795 axial slices, respectively, as shown in Table 1. Due to the inclusion of axial slices which did not contain BL in the study group, the total number of axial slices w/o BL was markedly higher than the total number of slices with BL.

Table 1

The details of the CBCT datasets used for training, validation and testing
		Training (N = 50)		Validation (N = 10)		Testing (N = 22)		Total (N = 82)
		BL	w/o BL	BL	w/o BL	BL	w/o BL	BL	w/o BL
Size of dataset	Number of cases	25	25	5	5	11	11	41	41
	Number of axial slices with BL	2754	0	396	0	1405	0	4555	0
	Number of axial slices without BL	7282	10178	1711	2423	1945	3445	10938	16046
Demographics	Female/Male	12:13	14:11	5:0	4:1	5:6	8:3	22:19	26:15
Demographics	Age (Mean ± SD, Range)	34 ± 19 (4–67)	61 ± 11.8 (41–79)	46 ± 21 (12–68)	41 ± 23.7 (19–74)	27 ± 15 (3–51)	50 ± 16 (26–70)	33 ± 18.9 (3–68)	56 ± 16 (19–79)
Technical details	Type of machine ICAT/MORITA/CRANEX	2:17:6	6:18:1	1:2:2	2:3:0	1:9:1	4:7:0	4:28:9	12:28:1
	Field of View Small/Medium/Large	12:9:4	5:10:10	2:2:1	1:2:2	6:3:2	2:5:4	20:14:7	8:17:16
	Range of voxel size (mm)	0.08–0.25	0.08–0.25	0.125-0.2	0.125-0.2	0.08–0.25	0.125–0.25	0.08–0.25	0.08–0.25
	Rotation 360⁰/180⁰	19:6	24:1	4:1	4:1	9:2	10:1	32:9	38:3
	Current (mA)	6.7 ± 2.5	5.8 ± 1.1	7 ± 3.4	5.2 ± 0.5	5.5 ± 1.1	5.5 ± 0.6	6.4 ± 2.4	5.6 ± 0.9
Radiographic features of Bone Lesion	Ant/PM/Molar/Ascend	10:6:6:3	NR	1:3:1:0	NR	1:3:6:1	NR	12:12:13:4	NR
	Follicular/Non-follicular	7:18	NR	1:4	NR	4:7	NR	12:29	NR
	Adjacent to IAN(%, No.)	52% (N = 13)	NR	40% (N = 2)	NR	82% (N = 9)	NR	58.5% (N = 24)	NR
	Minimal diameter (Mean ± SD, Range) (mm)	8.0 ± 3.6 (3–18)	NR	7.4 ± 1.7 (6–10)	NR	14 ± 7.4 (4–13)	NR	7.9 ± 3.4 (3–18)	NR
	Maximal diameter (Mean ± SD, Range) (mm)	24.0 ± 13.0 (7–63)	NR	13.4 ± 4.7 (8–20)	NR	25.4 ± 9.9 (12–41)	NR	23.1 ± 12.0 (7–63)	NR
	Histopathology of the lesion	RC(8), KOT(8), DC(5), AF(1), OF(1), CAm(1), NC(1)	NR	RC(3), KOT(1), LPC(1)	NR	KOT(4), RC(3), DC(1), AF(1), AOT(1), SBC(1)	NR	RC(14), KOT(13), DC(6), AF(2), OF(1), CAm(1), NC(1), LPC(1), AOT(1), SBC(1)	NR
Abbreviations: N = Number, BL = Bone lesion, w/o BL = without bone lesion ,IAN = Inferior Alveolar Nerve, RC = Radicular Cyst, KOT = Keratocystic Odontogenic Tumor, DC = Dentigerous Cyst, AF = Ameloblastic Fibroma, OF = Ossifying Fibroma, CAm = Cystic Ameloblastoma, NC = Nasopalatine Cyst, LPC = Lateral Periodontal Cyst, AOT = Adenomatoid Odontogenic Tumor, SBC = Simple Bone Cyst

All CBCT slices were resized to 800x800 pixels and converted from DICOM to JPEG format with 256 gray level values. The gray level values were automatically modified according to the window attributes in the DICOM header. The training and the validation datasets were augmented by horizontal flipping of the axial slices.

Table 1 summarizes the details of the CBCT datasets regarding demographics, scan parameters and BL characteristics. Data retrieved from the patient’s file included gender, age, and histopathological diagnosis. The CBCT scans were preformed using three different CBCT devices. Most scans were obtained with the MORITA™ 3D Accuitomo 170 followed by I-CAT™ Next Generation and CRANEX®3D Dental Imaging System. The tube voltage ranged between 85–120 kV, tube current ranged between 3–13 mA, field of view (FOV) diameter ranged between 4–16 cm, height ranged between 4–14 cm and the voxel size ranged between 0.08–0.25 mm. Most of the scans used full rotation, though some cases used half rotation. The number of the axial slices per case ranged between 161–705 (Appendices A and B).

BL cases included the following ten, histologically confirmed, pathologies: Keratocystic Odontogenic Tumor (KOT), Radicular Cyst (RC), Simple Bone Cyst (SBC), Ameloblastic Fibroma (AF), Ossifying Fibroma (OF), Adenomatoid Odontogenic Cyst (AOC), Dentigerous Cyst (DC), Cystic Ameloblastoma (CAm), Nasopalatine Cyst (NC) and Lateral Periodontal Cyst (LPC). Most of the lesions were mandibular (83%), adjacent to the Inferior Alveolar Nerve (IAN) (58.5%) and equally distributed between the anterior region (Ant), the premolar region (PM), and the molar region while lesser were in the ascending ramus (Ascend). In each case, the lesion location, contour and diameter were different in every axial slice. The maximal lesion diameter, defined as the lesion diameter in the slice in which it appears the largest, ranged between 7–63 mm (mean 23.07 ± 11.99 mm). The minimal lesion diameter, defined as the lesion diameter in the slice in which it appears the smallest, ranged between 3–18 mm (mean 7.85 ± 3.36 mm).

B. BONE LESION DETECTION

a. The deep learning Mask-RCNN algorithm

In the first step of the 3D detection process, a Mask RCNN deep learning algorithm (26) was implemented to detect the BL separately in each CBCT axial slice. This algorithm extends the Faster RCNN process (27) which detects only a bounding box containing the bone lesion, by additionally segmenting the bone lesion mask within the bounding box.

In order to train the Mask-RCNN algorithm, the exact contours of the BL in the training and the validation datasets were manually drawn in a middle axial slice and both the most inferior slice and the most superior slice by an Oral Maxillofacial radiologist (RAA). Based on these marks, the contours of the BL in all other slices in between, were manually drawn by students (TB, AIP, AA) using the VGG Image Annotator (VIA) program (28). Finally, these manual contours were confirmed, and modified when necessary, by an Oral Maxillofacial radiologist (CN( having more than 10 years' experience.

Mask RCNN algorithm uses a Feature Pyramid Network (FPN) backbone for feature extraction from each slice and then applies a Region Proposal Network (RPN) to propose Regions of Interest (ROIs) which may contain BL. The proposed ROIs are then resized to yield an aligned-similar size for all ROIs in order to classify them as a normal region or a region containing a bone lesion. Then, the algorithm generates a pre-segmentation bounding box (yellow rectangle in Fig. 1) for each detected bone lesion and adds an instant segmentation process (29) for producing a mask (red marking in Fig. 1) for the lesion within each bounding box (Fig. 1).

The ResNet 101 FPN backbone was selected for feature extraction with the initial weights obtained from pre-training on the Common Objects in Context (COCO) dataset (30). The Mask RCNN algorithm was trained with a single NVIDIA Quadro P2000 GPU, using the Stochastic Gradient Descent (SGD) optimizer algorithm for minimizing the loss of the model (31) with a momentum of 0.9, starting with a learning rate of 0.001 and ending with a learning rate of 0.0001. The training process ran for 96 hours, with 60 epochs and 6,000 steps per epoch. We found that fewer epochs led to underfitting and more epochs did not lead to a noticeable improvement. The minimum confidence level for categorizing an ROI as containing a bone lesion in a CBCT axial slice was set to 85%. A higher confidence level resulted in missing many relevant ROIs and a lower confidence level caused many false marks.

b. Improving the lesions detection by gray-level filtering

For improving the performance of the Mask-RCNN algorithm, an additional algorithm was used for filtering the detected ROIs by the mean gray level value of the segmented lesion within the ROI. ROIs containing a very bright segmented objects, with a mean gray level value higher than 155, were removed assuming that they represent teeth or restorations as shown in Fig. 2A. Similarly, ROIs containing a dark segmented object, with mean gray level value less than 50, were also removed assuming they represent air as shown in Fig. 2B.

c. Improving the detection of lesions by analyzing sequential axial slices

In order to confirm that the segmented object in each detected ROI in a specific slice truly represented a bone lesion, there was a need to analyze all sequential slices of each CBCT examination. For this purpose, an additional algorithm was developed to analyze all the slices of the CBCT examination and to evaluate in sequential slices the spatial location of the suspected objects (masks). For each slice containing a suspected object, the algorithm calculated its overlapping with suspected objects in previous slices by:

$$Overlapping=\frac{A\cap B}{A}$$

(1)

Where A is the number of pixels in the suspected object and $A\cap B$ is the number of pixels that are located identically in the suspected objects in this slice and in a previous slice, as illustrated in Fig. 3. We used several previous slices and not only a single slice, since the Mask RCNN algorithm may have missed the ROI containing the bone lesion in the neighboring slice, as shown in Fig. 4. Then, the algorithm defined subgroups of different ROIs in which the calculated overlapping between the objects was more than 50 percent. Any subgroup including at least 14 ROIs in any 20 successive slices was considered as a subgroup which contain a true bone lesion.

When the algorithm classified a CBCT case as containing a true bone lesion, it was important to identify the initial slice and the final slice containing the bone lesion. In the extreme slices the Mask RCNN may miss more ROIs since in these slices the lesion appears smaller and its borders are less defined. Therefore, the initial slice was defined as the first of at least 6 slices out of 20 successive slices, included in the same subgroup of ROIs. The final slice was defined as the last slice in the subgroup that fulfilled the same condition (6 out of 20 successive slices).

Then, the algorithm inserted a bone lesion mask in all the slices between the initial and final slices in which the bone lesion had been missed. This was performed by interpolating the shape of the detected objects in the two nearest neighboring slices since the mask of the bone lesion should be very similar in subsequent slices, as shown in Fig. 4. Finally, the algorithm removed all the suspected objects not included in the range of the initial and final slices.

Statistical analysis

The sensitivity of the algorithm for each CBCT case was defined as the ratio between the number of ROIs correctly detected and the total number of ROIs that actually contained a bone lesion. Precision was defined as the ratio between the number of ROIs correctly detected and the total number of ROIs detected by the algorithm, including false marks.

a. Detecting the bone lesion separately in each CBCT slice

The deep learning algorithm was tested on 22 cases included in the testing dataset (11 control cases and 11 cases with various types of benign BL). The BL differed by their size and location, some were adjacent to a tooth, cortical bone, the maxillary sinus or the nerve canal. The CBCT images of these cases were obtained with three CBCT devices using various FOVs, with different resolutions, tube parameters and rotation angles. The Mask RCNN algorithm detected, with high accuracy, the presence of all different types and forms of BL as well as their precise contour, independent of the image parameters, as shown in Fig. 5. However, the algorithm also missed some BL, as shown in Fig. 4 and sometimes falsely detected regions without lesions, such as regions of bone marrow defects, as shown in Fig. 6.

The performance of the Mask RCNN algorithm, for each of the 11 CBCT cases with BL, is summarized in Table 2. The columns marked by “Basic“ refer to the performance of the Mask RCNN algorithm on individual axial slices, before the analysis of sequential slices. The sensitivity of the deep learning algorithm ranged between 52.8%-100% with a total sensitivity of 83.5%. The precision of the algorithm for these cases ranged between 80.8%-100% with a total precision of 88.4%. For the 11 control cases in the testing dataset, which included 3445 axial slices altogether, the Mask RCNN algorithm falsely marked 112 objects, with an average of 10.2 false marks in all the slices of one case.

Table 2

**The algorithm performance**
Case No.	No. of ROIs containing lesions	No. of detected ROIs		No. of missed ROIs		No. of false marks		Sensitivity		Precision
Case No.	No. of ROIs containing lesions	Basic	Final	Basic	Final	Basic	Final	Basic	Final	Basic	Final
1	60	33	56	27	4	5	0	55.0%	93.3%	86.8%	100.0%
2	152	131	152	21	0	30	1	86.2%	100.0%	81.4%	99.3%
3	80	69	73	11	8	16	0	86.3%	91.3%	81.2%	100.0%
4	189	185	189	4	0	39	0	97.9%	100.0%	82.6%	100.0%
5	207	198	207	9	0	4	0	95.7%	100.0%	98.0%	100.0%
6	194	154	194	40	0	31	16	79.4%	100.0%	83.2%	92.4%
7	147	147	147	0	0	2	0	100.0%	100.0%	98.7%	100.0%
8	87	81	82	6	5	8	0	93.1%	94.3%	91.0%	100.0%
9	86	58	74	28	12	9	0	67.4%	86.0%	86.6%	100.0%
10	61	42	50	19	11	10	0	68.9%	82.0%	80.8%	100.0%
11	142	75	123	67	19	0	0	52.8%	86.6%	100.0%	100.0%
Total	1405	1173	1347	232	59	154	17	83.5%	95.9%	88.4%	98.8%

b. Classification of the CBCT cases following analysis of sequential slice

After the initial detection of suspected objects in each slice, all the slices of a specific CBCT scan were collectively analyzed, in order to evaluate if there is any subgroup of suspected objects that appears in sequential slices and represent a true bone lesion. The results of the sequential analysis showed that the all the 22 cases in the testing dataset were classified correctly with an accuracy of 100%. Namely, all 11 control cases were classified by the algorithm as cases w/o BL while the 11 CBCT scans with BL, were classified as containing BL.

c. 3D segmentation of the bone lesions following analysis of sequential slices

For all the CBCT cases which were classified as cases with BL, the initial and final slices including the lesion were automatically identified and the spatial location of the bone lesion in all the slices was calculated. Then, all the suspected objects that were segmented by the mask RCNN algorithm but did not overlap the spatial location of the bone lesion were removed. Finally, the algorithm segmented the BL in the slices in which they were missed, by interpolating the masks detected in the nearest neighboring slices. As a result, following the sequential analysis of the slices in each case, the improved sensitivity ranged between 82.0%-100% with a total sensitivity of 95.9%. and the improved precision ranged between 92.4%-100% with a total precision of 98.8%, The improved performance is shown in Table 2, in the columns marked by “Final”.

Following the final segmentation of the BL in the relevant CBCT slices, the algorithm generated a 3D segmentation of each bone lesion, as shown in Fig. 7. The volume of the bone lesion, which is important for follow-up, is defined as the product of the voxel volume by the number of the segmented voxels in all the slices. The total time for detecting and segmenting the BL in each CBCT case was between 12–15 minutes, depending on the total number of axial slices analyzed.

In this study an automated Mask-RCNN based deep leaning algorithm was used for the detection and 3D segmentation of BL in maxillofacial CBCT scans. The algorithm was trained, validated and tested on 82 CBCT scans with and without histologically-confirmed benign BL, using 3 different CBCT devices with various protocols. The BL were detected in individual axial images with high sensitivity (95.9%) and high precision (98.9%). An accuracy of 100% was yielded for categorizing each CBCT case as containing BL or not.

Various methods have been developed for detecting and classifying BL in the maxillofacial region, and specifically in CBCT. Automatic detection of BL in CBCT is critical since sometimes the lesions may be located outside the region of interest, irrelevant to the referral reason and therefore may be missed (32). Setzer et al. (20) used CNN for segmenting periapical lesions, following manual cropping of a limited CBCT volume including the tooth and the adjacent lesion with a sensitivity of 88% and precision of 87%. However, this algorithm cannot be used for detecting incidental BL. Orhan et al. (22) developed a CNN algorithm which automatically generated a 3D box around each tooth, and then segmented in the relevant boxes any periapical lesion with a sensitivity of 89% and precision of 95%. Two studies of Abdolali et al.(23, 24) dealt with automatic detection of benign BL in various locations in CBCT images. Their algorithm for 3D segmentation and classification of three bone lesion types (RC, DC and KOT) resulted in a total classification rate of 96.5%. However, the method used in these studies requires a medium to large CBCT field-of-view since it is based on asymmetry between both sides of the jaw. Moreover, this method might not be suitable for cases with a solitary anterior lesion, or of multiple symmetrical lesions.

In this study, we developed a fully-automated deep learning algorithm for the detection and 3D segmentation of maxillofacial benign BL in CBCT imaging, independent of their size, shape, location or histopathology. This algorithm automatically detects the location of the lesions; it is not based on the symmetry of the jaw, nor is it limited to periapical lesions only. Moreover, this algorithm is most appropriate for detecting incidental findings, as it was trained and tested on a clinically realistic dataset which was acquired by various CBCT devices, using different protocols.

The study included 41 cases with a variety of histologically confirmed benign BL and 41 cases without BL, scanned by one of three CBCT devices with a variety of protocols. Our dataset, retrieved from our archives during the last two years, is a representative of the different types of bone lesions, we routinely encounter, in our maxillofacial imaging unit. It should be noted that since this study was conducted in a hospital based tertiary center, the prevalence of these BL in the archives, is much higher than in community-based imaging centers. In addition, it is customary to train and test the deep learning algorithms on equal number of cases in each category and therefore we selected identical number of cases with BL and without BL although it does not represent their overall prevalence.

The developed algorithm included a deep learning stage, using Mask-RCNN architecture, followed by gray level filtration and analysis of sequential slices. Even though the BL were diverse and the CBCT scans were obtained using different devices and a variety of imaging protocols, the algorithm managed to accurately distinguish between CBCT cases with and without BL with an accuracy of 100%. Furthermore, the algorithm detected the BL in individual CBCT axial slices with high sensitivity (95.9%) and high precision (98.8%). Moreover, the algorithm segmented only the BL and distinguished between the lesion and other hypodense anatomical structures, such as the maxillary sinus or the inferior alveolar nerve, which were adjacent to some of the lesions (Fig. 5). Mask RCNN network was trained to detect and segment the BL in each CBCT slice individually, resulting in thousands of CBCT slices for the deep learning training process. Moreover, the number of CBCT slices in the dataset was doubled by horizontal flipping. Increasing the dataset with more cases of each lesion type, may allow the deep learning algorithm to classify the type of the lesion (17–19) although they look very similar to the human eye.

In contrast to Multi Detector CT scans, in CBCT, the HU units are not reproducible and depend on the CBCT device that were used (33) .The images were converted from DICOM to JPEG and gray levels were used instead of HU, since in order to yield good detection results, we used transfer learning on DL architecture that has been already trained on images with 256 gray level values. This automatic conversion is quick and does not significantly affect the calculation time. In the final stage, the original pixel dimensions of the DICOM images were used to reconstruct the three-dimensional model.

Our study focused on 10 types of hypodense benign BL with similar gray level values and internal structures appearances in CBCT. Our algorithm successfully segmented the BL volumetrically. The lesion’s volume is not crucial for diagnosis, but it may serve in treatment planning and follow-up, in cases such as simple bone cyst (34), or following marsupialization of KOT (35). Furthermore, the 3D segmentation of BL may assist in virtual reality educational simulators which are increasingly being used, for both diagnosis and treatment planning, especially now in the COVID-19 era (36). The relatively long time required for running the algorithm on each CBCT case, can be resolved by using parallel computing with more advanced GPUs.

Our novel algorithm could be used for improving the capability of CBCT scans interpretation by clinicians. Though according to international guidelines CBCT scans interpretation is mandatory, it is still not a common practice worldwide (37). Thus, increasing the detection sensitivity of incidental bone lesions in early stages will improve patient management. In addition, this algorithm should also improve the diagnostic results of less experienced clinicians.

The major limitation of our study its constriction on pathologies with similar radiographic features. Thus, we only included in this study benign BL which usually have well defined borders. Future studies should include malignant BL with less defined borders which are clinically important, especially if overlooked (32). Another limitation of the study is that our testing dataset has relatively a limited number of cases. In addition, the retrospective nature of our dataset collection does not reflect the performance of our algorithm on cases from community based-imaging centers since the ratio of cases with and without BL does not represent the BL incidence in overall CBCT scans. Therefore, a prospective, multi-centered study, with larger datasets, from several community-based imaging centers, would be appropriate to examine the algorithm practical role in patient care.

We developed a unique deep learning algorithm for detecting a variety of benign BL in varied volumetric maxillofacial CBCT imaging. Reliable detection of incidental findings is important for BL early diagnosis. Automated detection may monitor BL treatment in CBCT follow-up imaging. The 3D segmentation of BL may be used for educational purposes. Future work should include a larger study group with mixed and hyperdense lesions as well as malignant lesions.

All authors declare no financial and non-financial conflicts of interest.

This study was approved by the Institutional Review Board (HMO-0297-21). This study was retrospective and involved only medical record analysis, therefore, the IRB approved an exemption from informed consent.

No funding was obtained for this study

Authors contribution TY, RAA IL and CN Initiated the research, TY, IL and CN formed the methodology, ABN, RAA, TB, AIP and AA collected and segmented the data, TY and SLD did the software formal analysis, RAA and CN created the original draft, RAA, CN, TY, AC and IL reviewed and edited the manuscript, TY and CN coordinated the research

Dawood A, Patel S, Brown J. Cone beam CT in dental practice. Br Dent J. 2009;
Drage N, Rogers S, Greenall C, Playle R. Incidental findings on cone beam computed tomography in orthodontic patients. J Orthod. 2013;
Lopes IA, Tucunduva RMA, Handem RH, Capelozza ALA. Study of the frequency and location of incidental findings of the maxillofacial region in different fields of view in CBCT scans. Dentomaxillofacial Radiol. 2016;
Warhekar S, Nagarajappa S, Dasar PL, Warhekar AM, Parihar A, Phulambrikar T, et al. Incidental findings on cone beam computed tomography and reasons for referral by dental practitioners in Indore city (M.P). J Clin Diagnostic Res. 2015;
Allareddy V, Vincent SD, Hellstein JW, Qian F, Smoker WRK, Ruprecht A. Incidental findings on cone beam computed tomography images. Int J Dent. 2012;
Fiaschetti V, Fanucci E, Rascioni M, Ottria L, Barlattani A, Simonetti G. Jaw expansive lesions: population incidence and CT dentalscan role. Oral Implantol (Rome). 2010;
Flaitz CM, Hicks J. Delayed tooth eruption associated with an ameloblastic fibro-odontoma. Pediatr Dent. 2001;
Araújo JP, Kowalski LP, Rodrigues ML, de Almeida OP, Lopes Pinto CA, Alves FA. Malignant Transformation of an Odontogenic Cyst in a Period of 10 Years. Case Rep Dent. 2014;
Scarfe WC, Toghyani S, Azevedo B. Imaging of Benign Odontogenic Lesions. Radiologic Clinics of North America. 2018.
Heo M-S, Kim J-E, Hwang J-J, Han S-S, Kim J-S, Yi W-J, et al. Artificial intelligence in oral and maxillofacial radiology: what is currently possible? Dentomaxillofacial Radiol. 2021;
McBee MP, Awan OA, Colucci AT, Ghobadi CW, Kadom N, Kansagra AP, et al. Deep Learning in Radiology. Academic Radiology. 2018.
Hwang JJ, Jung YH, Cho BH, Heo MS. An overview of deep learning in the field of dentistry. Imaging Sci Dent. 2019;
Hung K, Montalvao C, Tanaka R, Kawai T, Bornstein MM. The use and performance of artificial intelligence applications in dental and maxillofacial radiology: A systematic review. Dentomaxillofacial Radiol. 2019;
Lahoud P, EzEldeen M, Beznik T, Willems H, Leite A, Van Gerven A, et al. Artificial Intelligence for Fast and Accurate 3-Dimensional Tooth Segmentation on Cone-beam Computed Tomography. J Endod. 2021;
Chen Y, Du H, Yun Z, Yang S, Dai Z, Zhong L, et al. Automatic Segmentation of Individual Tooth in Dental CBCT Images from Tooth Surface Map by a Multi-Task FCN. IEEE Access. 2020;
Jang TJ, Kim KC, Cho HC, Seo JK. A fully automated method for 3D individual tooth identification and segmentation in dental CBCT. IEEE Trans Pattern Anal Mach Intell. 2021;
Yilmaz E, Kayikcioglu T, Kayipmaz S. Computer-aided diagnosis of periapical cyst and keratocystic odontogenic tumor on cone beam computed tomography. Comput Methods Programs Biomed. 2017;
Flores A, Rysavy S, Enciso R, Okada K. Non-invasive differential diagnosis of dental periapical lesions in cone-beam CT. In: Proceedings – 2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, ISBI 2009. 2009.
Lee JH, Kim DH, Jeong SN. Diagnosis of cystic lesions using panoramic and cone beam computed tomographic images based on deep learning neural network. Oral Dis. 2020;
Setzer FC, Shi KJ, Zhang Z, Yan H, Yoon H, Mupparapu M, et al. Artificial Intelligence for the Computer-aided Detection of Periapical Lesions in Cone-beam Computed Tomographic Images. J Endod. 2020;
Huang J, Yan H, Li J, Stewart HM, Setzer F. Combining anatomical constraints and deep learning for 3-D CBCT dental image multi-label segmentation. In: Proceedings - International Conference on Data Engineering. 2021.
Orhan K, Bayrakdar IS, Ezhov M, Kravtsov A, Özyürek T. Evaluation of artificial intelligence for detecting periapical pathosis on cone-beam computed tomography scans. Int Endod J. 2020;
Abdolali F, Zoroofi RA, Otake Y, Sato Y. Automatic segmentation of maxillofacial cysts in cone beam CT images. Comput Biol Med. 2016;
Abdolali F, Zoroofi RA, Otake Y, Sato Y. Automated classification of maxillofacial cysts in cone beam CT images using contourlet transformation and Spherical Harmonics. Comput Methods Programs Biomed. 2017;
Pauwels R, Araki K, Siewerdsen JH, Thongvigitmanee SS. Technical aspects of dental CBCT: State of the art. Dentomaxillofacial Radiology. 2015.
Proceedings of the IEEE International Conference on Computer Vision. Proceedings of the IEEE International Conference on Computer Vision. 2017.
Ren S, He K, Girshick R, Sun J. Faster R-CNN: Towards real-time object detection with region proposal networks. In: Advances in Neural Information Processing Systems. 2015.
Dutta A, Zisserman A. The VIA annotation software for images, audio and video. In: MM 2019 - Proceedings of the 27th ACM International Conference on Multimedia. 2019.
Almubarak H, Bazi Y, Alajlan N. Two-stage mask-RCNN approach for detecting and segmenting the optic nerve head, optic disc, and optic cup in fundus images. Appl Sci. 2020;
Lin TY, Maire M, Belongie S, Hays J, Perona P, Ramanan D, et al. Microsoft COCO: Common objects in context. In: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). 2014.
Ketkar N. Stochastic gradient descent. In: Deep learning with Python. Berkeley, CA: Apress; 2017. p. 113–32.
Abdalla-Aslan R, Friedlander-Barenboim S, Aframian DJ, Maly A, Nadler C. Ameloblastoma incidentally detected in cone-beam computed tomography sialography: A case report and review of the literature. J Am Dent Assoc. 2018;
Razi T, Emamverdizadeh P, Nilavar N, Razi S. Comparison of the Hounsfield Unit in CT scan with the Gray Level in cone-beam CT. J Dent Res Dent Clin Dent Prospects. 2019;
Sapp JP, Stark ML. Self-healing traumatic bone cysts. Oral Surgery, Oral Med Oral Pathol. 1990;
Queiroz IV. Nonsurgical Management of Keratocystic Odontogenic Tumor: Report of a Clinical Case. J Oral Hyg Heal. 2013;
Pohlenz P, Gröbe A, Petersik A, Von Sternberg N, Pflesser B, Pommert A, et al. Virtual dental surgery as a new educational tool in dental school. J Cranio-Maxillofacial Surg. 2010;
Brown J, Jacobs R, Levring Jäghagen E, Lindh C, Baksi G, Schulze D, et al. Basic training requirements for the use of dental CBCT by dentists: A position paper prepared by the European Academy of Dento Maxillo Facial Radiology. Dentomaxillofacial Radiol. 2014;

No competing interests reported.

PDF

Reviewers invited by journal
01 May, 2022
Editor assigned by journal
01 May, 2022
Submission checks completed at journal
27 Apr, 2022
First submitted to journal
17 Apr, 2022

You are reading this latest preprint version

Deep learning algorithm for the detection and 3D segmentation of maxillofacial bone lesions in CBCT

Status:

Version 1

Abstract

Figures

Introduction

Materials And Methods

A. THE DATASET

a. The deep learning Mask-RCNN algorithm

b. Improving the lesions detection by gray-level filtering

c. Improving the detection of lesions by analyzing sequential axial slices

Statistical analysis

Results

a. Detecting the bone lesion separately in each CBCT slice

Discussion

Conclusions

Declarations

References

Competing Interests

Supplementary Files

Status:

Version 1