Primary clinical study of radiomics for diagnosing simple bone cyst of the jaw

Abstract

Objective:

To screen the radiomic features of simple bone cysts of the jaws and explore the potential application of radiomics in pre-operative diagnosis of jaw simple bone cysts.

Methods:

The investigators designed and implemented a case–control study. 19 patients with simple bone cysts who were admitted to the Department of Maxillofacial Surgery, Sun Yat-sen University Affiliated Stomatology Hospital from 2013 to 2019 were included in this study. Their clinical data and cone-beam computed tomography (CBCT) images were examined. The control group consisted of patients with odontogenic keratocyst. CBCT imaging features were analyzed and compared between the patient and control groups.

Results:

Overall, 10,323 image features were extracted through feature analysis. A subset of 25 radiomic features obtained after feature selection were analyzed further. These 25 features were significantly different between the 2 groups (p < 0.05). The absolute value of correlation coefficient was 0.487–0.775. Gray-level co-occurrence matrix (GLCM) contrast, neighborhood gray tone difference matrix (NGTDM) contrast, and GLCM variance were the features with the highest correlation coefficients.

Conclusions:

Pre-operative radiomics analysis showed the differences between simple bone cysts and odontogenic keratocysts, can help to diagnose simple bone cysts. Three specific texture features—GLCM contrast, NGTDM contrast, and GLCM variance—may be the characteristic imaging features of simple bone cysts of the jaw.

Introduction

Simple bone cyst (SBC) of the jaw is a rare non-odontogenic cyst in the oral cavity, accounting for about 1% of all oral cysts.¹ It is a non-neoplastic bone lesion without an epithelial lining and soft tissue, which is empty or filled with fluid, and without evidence of infection.² SBC usually progresses slowly or does not progress, without swelling, pain, and other functional signs. Therefore, we often diagnose SBC incidentally during imaging examinations.³ Majority of SBCs show homogeneous osteolysis surrounded in a single cavity on panoramic radiograph, surrounded by narrow conical bony condensation. Some cases present multilocular with septum-like images, can be misdiagnosed with other maxillary cystic lesions such as odontogenic keratocyst (OKC).⁴ OKC grows through the bone marrow tissue without causing expansion, resulting in a similar imaging features with SBC. Due to the aggressive behavior and high recurrence rate, OKC requires more radically surgical treatment, such as curettage, mandibular segmental resection and so on. The diagnosis of OKC depended on microscopic examination, while SBCs need to be confirmed via surgical exploration.⁵ Since 1955, surgical intervention has been the first choice for the diagnosis and treatment of SBC.^6,7 An in-depth study of SBC showed that SBC can resolve spontaneously.⁸ If we can diagnose SBC accurately before surgery, then conservative treatment can be the first choice. An invasive surgery can be avoided, if the condition of the cyst needs to be assessed regularly. However, there are no effective pre-operative diagnostic methods. Thus, an accurate non-invasive approach that can diagnose SBC pre-operatively is urgently needed to help surgeons make precise treatment decisions.

Recently, there has been an increasing interest in extracting quantitative features from medical images to improve clinical diagnosis and treatment. Radiomics is a new medical frontier based on the idea that microscopic gene or protein pattern changes are expressed on macroscopic images, and through in-depth mining of image features, it can identify the changes of human tissue, cell and gene level. Specifically, radiomics mainly focuses on improving image analysis by using high-throughput extraction of large amounts of features from radiographic images.⁹ For example, the features of chest X-ray images have been applied in the diagnosis of pediatric pneumonia and pulmonary tuberculosis.¹⁰ Radiomics is also used to predict the histologic grade of oral squamous cell carcinoma.¹¹ Although radiomics has not been applied in jaw diseases, it may be used as an accurate non-invasive approach to diagnose SBC.

The present study aims to explore the application of radiomics in distinguish SBC and other maxillary cystic lesions (OKC). Extracting the radiomics features between SBC and OKC, and identifying the specific features of SBC, which are helpful to the the pre-operatively accurate diagnosis of SBC.

Patients and methods

Patients selection

A retrospective, case–control study was conducted to address the research purpose. The study was approved by the Institutional Ethics Board of the Hospital of Stomatology, Sun Yat-sen University, and patient informed consent was waived. In the present study, we enrolled 38 patients, including 19 patients with SBC and 19 with OKC, who had completed treatment between 2013 and 2019. All patients with OKC were diagnosed by pathological examination according to the World Health Organization histologic classification.² The OKC patients with nevoid basal cell carcinoma syndrome were excluded in the study. The diagnosis of SBC was confirmed by surgery exploration based on the presence of a cavity, which may contain blood or blood-carrying fluid, and surrounded by a hard capsule.¹² The clinical and imaging data of those patients were retrospectively retrieved from the archive of Hospital of Stomatology, Sun Yat-sen University. The following data were collected: (1) clinical data: gender, age, pain, bulging of the jaw, and history of trauma; (2) imaging data: pre-operative panoramic radiographs and CBCT images. All panoramic radiography images were taken on the same Promax Digital Panoramic X-ray Machine (Planmeca, Finland). The following imaging parameters were used: resolution ratio 2500*1244 pixels, Grayscale 4096, grayscale resolution 12 bits, panoramic image area ≥15*30 cm, image amplification factor 1.27 times. All CBCT images were obtained using a CBCT scanner (DCTPRO, VATECH, Yongin-Si, Republic of Korea) with a field of view of 16 × 7 cm and a voxel size of 0.16 mm. The operating parameters were set at 90.0 kV and 9 mA with a scanning time of 24 s.

Methods and materials

Extraction of the region of interest (ROI) area

The CBCT data of patients with SBC and OKC were imported into the image processing software, ITK-SNAP 3.2 (www.itksnap.org).¹³ ITK-SNAP program is a free, open-source, and multiplatform image analysis tool. It provides semi-automatic segmentation using active contour methods, as well as manual delineation and image navigation and supports for many different 3D image formats, including NIfTI and DICOM. One dentomaxillofacial radiologist then manually outlined and segmented the area of the cystic lesions along their edges, layer by layer, to generate ROIs. The CBCT image and the corresponding segmentation results are shown in Figure 1.

Figure 1.

ROI segmentation results from CBCT images. (A) A ROI was manually outlined from CBCT, (B) Axial plane, (C) sagittal plane and (D) coronal plane. The cyst is marked in red. CBCT, cone-beam computed tomography; ROI, region of interest.

Feature extraction

In the present study, we computed the extracted 43 types of texture features using 240 different combinations of extraction parameters. These texture features consisted of 3 first-order statistical features, 9 gray-level co-occurrence matrix (GLCM) features, 13 gray-level run-length matrix (GLRLM) features, 13 gray-level size zone matrix (GLSZM) features, and 5 neighborhood gray tone difference matrix (NGTDM) features (Table 1). The four following extracted parameters were applied for extracting texture features: wavelet band-pass filtering, isotropic resampling, grayscale quantization algorithm, and quantization level.¹⁴ Their corresponding values are: (1/2, 2/3, 1, 3/2, 2), [pixelW (Original image resolution), 1 mm, 2 mm, 3 mm, 4 mm, 5 mm], (Equal-probability, Lloyd-Max) and (8, 16, 32, 64). Based on their corresponding values (n = 5, 6, 2, 4), 1,0324 ($4+(3+9+13+13+5)\times 5\times 6\times 2\times 4=10324$) features were extracted.

Table 1.

Characteristics of the imaging group used in this study

Feature type	Feature name
Non-texture	Volume
	Size
	Solidity
	Eccentricity
Texture
First-order statistics feature	Variance
	Skewness
	Kurtosis
GLCM (Gray-level co-occurence matrix)	Energy
	Contrast
	Correlation
	Sum-average
	Homogeneity
	Entropy
	Variance
	Dissimilarity
	Auto-Correlation
GLRLM (Gray-level run-length matrix)	Short Run Emphasis(SRE)
	Long Run Emphasis(LRE)
	Gray-Level Non-uniformity(GLN)
	Run-Length Non-uniformity(RLN)
	Gray-Level Variance(GLV)
	Run-Length Variance(RLV)
	Run Percentage(RP)
	Low Gray-Level Run Emphasis(LGRE)
	High Gray-Level Run Emphasis(HGRE)
	Short Run Low Gray-Level Emphasis(SRLGE)
	Short Run High Gray-Level Emphasis(SRHGE)
	Long Run Low Gray-Level Emphasis(LRLGE)
	Long Run High Gray-Level Emphasis(LRHGE)
GLSZM (Gray-level run-length matrix)	Small Zone Emphasis(SZE)
	Large Zone Emphasis(LZE)
	Gray-Level Non-uniformity(GLN)
	Zone-Size Non-uniformity(ZSN)
	Gray-Level Variance(GLV)
	Zone-Size Variance(ZSV)
	Zone Percentage(ZP)
	Low Gray-Level Zone Emphasis(LGZE)
	High Gray-Level Zone Emphasis(HGZE)
	Small Zone Low Gray-Level Emphasis(SZLGE)
	Small Zone High Gray-Level Emphasis(SZHGE)
	Large Zone Low Gray-Level Emphasis(LZLGE)
	Large Zone Low Gray-Level Emphasis(LZHGE)
NGTDM (Neighborhood gray tone difference matrix)	Coarseness
	Contrast
	Busyness
	Complexity
	Strength

Feature selection

To screen for features that correlate with SBCs, we used the sequence forward selection (SFS) method¹⁵ to optimize the useful features and filter the redundant ones. A feature subset containing 25 features with the highest correlation was then selected.

Statistical analyses

All statistical analyses were performed using the IBM SPSS 20.0 (SPSS Inc., Chicago, IL) statistical software. Spearman rank correlation analysis was performed on the data of each feature to compare the correlation between the features and categories. p values < 0.05 were regarded as indicating statistical significance.

Results

General case information

Clinical data of the SBC and OKC groups are shown in Table 2. In the present study, there were 19 patients with SBC with an average age of 17.0 years, of which 8 (42.1%) were males and 11 (57.9) were females. Six (31.6%) patients with SBC showed mandibular expansion, while the other patients had no symptoms. There were 19 patients with OKC with an average age of 29.1 years, of which 7 were males and 12 were females. Compared to patients with SBC, patients with OKC had a higher incidence of mandibular expansion; 17 (89.5%) patients with OKC had mandibular expansion. Moreover, unlike patients with SBC, seven patients with OKC experienced pain and four patients had tooth mobility.

Table 2.

Clinical data of the SBC and OKC groups

Group	Male	Female	Average age	Pain	Mandibular expansion	Tooth mobility	Past trauma
SBC	8	11	17.0	0	6	0	0
OKC	7	12	29.1	7	17	4	0

OKC, Odontogenic keratocyst;SBC, Simple bone cyst.

Imaging performance

Panoramic radiography

The average maximum diameter of SBC was 28.08 mm (Table 3). Of the 19 cases of SBC, most of them originated from the posterior teeth and ascending branch area of the mandible (73.7%), and a small portion of the cases occurred in the anterior teeth area (26.3%). Eight cases (42.1%) of SBC appeared cone-shaped, and the rest of the cases had an irregular shape. 15 patients (78.9%) with SBC showed scalloping on the panoramic radiograph; the upper boundary of the cyst extended till the root of the tooth, and it was fan-shaped with hard bone lines and resembled the shape of a shell (Figure 2). The pagination notches on the lower margin was observed in five cases. All patients with SBC did not have teeth displaced and roots absorbed.

Table 3.

Imaging findings of the SBC and OKC groups

Variables	SBC	OKC
Location, n (%)
Anterior teeth region	5 (26.3)	1 (5.3)
Posterior teeth region	12 (63.2)	13 (68.4)
Mandibular ramus region	2 (10.5)	5 (26.3)
Average size, (mm)	28.08	47.06
Cone shape, n (%)
Yes	8 (42.1)	3 (15.8)
No	11 (57.9)	16 (84.2)
Scalloping, n (%)
Yes	15 (78.9)	7 (36.8)
No	4 (21.1)	12 (63.2)
Pagination notch
Yes	5 (26.3)	9 (47.4)
No	14 (73.7)	10 (52.6)
Tooth displacement, n (%)
Yes	0 (0)	4 (21.1)
No	19 (100)	15 (78.9)
Tooth root resorption, n (%)
Yes	0 (0)	5 (26.3)
No	19 (100)	14 (73.7)
Cortical bone perforation, n (%)
Yes	0 (0)	7 (36.8)
No	19 (100)	12 (63.2)
Cortical bone thinning, n (%)
Yes	17 (89.5)	17 (89.5)
No	2 (10.5)	2 (10.5)
Relationships between cysts and the mandibular canal, n (%)
Cyst away from the canal	6 (31.6)	3 (15.8)
Cyst squeezed the canal	6 (31.6)	7 (36.8)
Cyst destroyed the bone around the canal	0 (0)	9 (47.4)
Bone ridges protruding into the cavity, n (%)
Yes	14 (73.7)	2 (10.5)
No	5 (26.3)	17 (89.5)

SBC, Simple bone cyst; OKC, odontogenic keratocyst

Figure 2.

Panoramic radiograph revealing the scalloping at the root of 47 (arrow).

Of the 19 patients with OKC, the average maximum diameter was 47.06 mm, which was larger than SBC. Similar to SBCs, 18 OKCs (94.7%) originated from the posterior teeth and ascending branch area of the mandible, and only 1 case (5.3%) occurred in the anterior teeth. Seven cases (36.8%) of OKC presented scalloping. Nine cases (47.4%) showed pagination notches on the lower margin and five cases showed root resorption.

CBCT

Via CBCT, we observed cortical bone perforation in 36.8% of OKCs and bone thinning in 89.5% of cases; in SBC cases, there was only cortical bone thinning and no perforation (Figure 3a). Three relationships between cysts and the mandibular canal could be found: (1) 31.6% of SBCs and 15.8% of OKCs were located far away from the mandibular canal; (2) 31.6% of SBCs and 36.8% of OKCs squeezed the canal; (3) 47.4% of OKCs destroyed the bone around the canal (Figure 4), but this did not occur in any of the SBC cases. Moreover, in SBC cases, most cysts (73.7%) showed bony ridges protruding into the lumen at the edge of the cyst wall (Figure 3b). However, in OKC cases, we observed this characteristic only in two cases (Table 3).

Figure 3.

CBCT manifestation of simple bone cyst. (A) A CBCT scan showed only a thinning cortical bone between simple bone cyst and mandibular canal, (B) A CBCT scan showed a bony ridge protruding into the lumen at the edge of the cyst wall. CBCT, cone-beam computed tomography.

Figure 4.

CBCT manifestation of OKC. A CBCT scan showed part of the bone around the mandibular canal is destroyed by the cyst. (A) sagittal plane, (B) coronal plane, (C) Axial plane. CBCT, cone-beam computed tomography; OKC, odontogenic keratocyst.

Results of radiomic feature selection

As a result of feature extraction, a total of 10 324 radiomic features were obtained from ROI, which consist of non-texture features and texture features. Through feature selection, we obtained a feature subset containing 23 texture features and 2 non-texture features. Table 4 presented Spearman’s rank correlation (Rs) and the corresponding p-value of 25 features in the feature subset. The two non-texture features (volume and size) were significantly different between the OKC and SBC groups. Among the texture features, GLCM contrast, NGTDM contrast, and GLCM variance had the highest correlation coefficients.

Table 4.

The results of correlation tests of 25 features in the feature subset

Features	R	Scale	Algorithm	Ng	Rs	p value
GLCM contrast	1.00	3	Lloyd	16	0.775	<0.0001
GLCM variance	0.50	4	Lloyd	8	0.706	<0.0001
GLCM dissimilarity	0.50	4	Lloyd	64	0.692	<0.0001
GLCM homogeneity	0.67	4	Lloyd	8	−0.567	0.000312
GLCM energy	2.00	5	Equal	16	0.631	<0.0001
NGTDM contrast	0.50	4	Lloyd	8	0.713	<0.0001
NGTDM strength	0.67	1	Lloyd	16	0.647	<0.0001
NGTDM coarseness	2.00	2	Equal	64	0.636	<0.0001
NGTDM busyness	2.00	2	Equal	8	−0.620	<0.0001
GLSZM GLV	0.67	2	Lloyd	16	0.690	<0.0001
GLSZM ZSN	0.50	4	Lloyd	8	0.653	<0.0001
GLSZM SZE	1.50	5	Equal	32	0.650	<0.0001
GLSZM ZP	1.00	5	Equal	8	0.639	<0.0001
GLSZM LZE	0.50	5	Equal	8	−0.638	<0.0001
GLSZM ZSV	1.50	5	Equal	64	0.634	<0.0001
GLRLM SRE	1.50	5	Equal	32	0.672	<0.0001
GLRLM RLN	1.50	5	Equal	16	0.666	<0.0001
GLRLM LRE	1.00	5	Equal	8	−0.664	<0.0001
GLRLM RLV	2.00	5	Lloyd	64	0.636	<0.0001
GLRLM RP	1.50	5	Equal	32	0.635	<0.0001
GLRLM GLN	2.00	5	Equal	64	0.629	<0.0001
Global variance	0.50	4	Equal	8	0.652	<0.0001
Global kurtosis	1.00	5	Equal	16	−0.556	0.000428
Volume	–	–	–	–	−0.540	0.000673
Size	–	–	–	–	−0.487	0.002615

GLCM, Gray-level co-occurrence matrix; GLRLM GLN, Gray-level run-length matrix gray-level non-uniformity; GLSZM GLV, Gray-level size zone matrix gray-level variance; GLRLM LRE, gray-level run-length matrix Long run emphasis; GLSZM LZE, Gray-level size zone matrix Large zone emphasis; NGTDM, Neighborhood gray tone difference matrix; GLRLM RLN, gray-level run-length matrix Run-length non-uniformity; GLRLM RLV, gray-level run-length matrix Run-length variance; GLRLM RP, gray-level run-length matrix Run percentage; Rs, Spearman rank correlation coefficients; GLRLM SRE, Gray-level run-length matrix Short run emphasis; GLSZM SZE, gray-level size zone matrix Small zone emphasis; GLSZM ZP, gray-level size zone matrix Zone percentage; GLSZM ZSN, gray-level size zone matrix Zone-size variance.

Discussion

SBC is a pseudocyst without a lining epithelium and often needs to be distinguished from true cysts because different treatments are required for these conditions.¹⁶ However, some unilocular and multilocular SBCs are difficult to distinguish from true cysts. In the present study, we compared the panoramic radiograph and CBCT scans of 19 SBCs and 19 OKCs to identify the diagnostic radiographic findings. There is a great difference in radiographic findings between SBCs and OKCs. We then extracted 10,324 features from every CBCT image in the SBC and OKC groups. Finally, 2 non-texture features and 23 texture features that had the distinguishable power were obtained by feature selection.

In the present study, some radiographic features of SBC were found. Scalloping is one of the typical features, wherein the upper boundary of the cyst extends until the root of the tooth, thereby having the shape of a fan. A previous study reviewed 1253 patients with SBC, about 40% of whom showed scalloping.¹⁷ Similarly, Sabino-Bezerra¹⁸ et al described six patients with SBC, three of whom had marginal scallop-like lesions. In our study, we observed scalloping on the panoramic radiograph of 15 patients with SBC (78.9%). It may be related to the epidemiological differences between different regions or the small sample size. Interestingly, Damante et al¹⁹ reported that in 10 cases of SBC with scalloping, 6 cases remodeled and 1 case spontaneously resolved. It suggested that the scalloping may be related to the prognosis of SBC. However, 36.8% patients with OKC showed scalloping on the panoramic radiographs in the present study. A previous study reported that the scalloping was more frequently observed in the posterior segment of OKCs than in SBCs on the panoramic radiographs.¹⁹ Therefore, the radiographic features of SBC are not diagnostic and may be confused with a variety of odontogenic cysts. CBCT was also applied for three-dimensional visualization of SBCs. Using CBCT, we observed that SBC showed a lower boundary that was flat and there were high-density irregular osteoporotic images or bony crests protruding into the lumen at the margin of the capsule wall. However, CBCT also cannot provide enough diagnostic information. The radiomic features consisted of the information about mandibular cystic lesions’ gray shadow, shape, intensity, texture, as well as spatial relationship. According to the radiomics hypothesis, the genomic heterogeneity may translate to expression in a disease intra heterogeneity that can be assessed through imaging.²⁰ With regard to gross observation and histopathological examination, the tissue composition of lesions from jaw SBC and OKC differs greatly. Maxillary SBCs may be empty or blood/fluid-filled and occasionally have a fibrous capsule wall, with the bone crest protruding into the cavity, or bone fragments scattered into the fluid. The cyst sac of OKCs have fluid and keratosis, and the sac wall thickens during infection. Therefore, radiomic features can reflect the difference between OKC and SBC.

In the present study, 10,324 texture features were extracted, including information such as gray shadow, intensity, texture, and spatial relationship of cystic lesions. According to the correlation between features and categories, some feature organization methods are applied to remove irrelevant, redundant, and noisy information from the data, and select the optimal features to further improve the differentiation performance.¹⁵ 25 features that have the ability to distinguish SBC and OKC were identified.

Two non-texture features, namely, volume and size were identified to have the ability to discriminate between SBC and OKC. Volume refers to the product of the number of pixels extracted and the dimension of the pixel, and size refers to the maximum diameter of the lesion area extracted from the CBCT image.²¹ Both these features correlated negatively, indicating that the larger the range of the maxillary cystic lesion, the lower the probability of it being an SBC. However, when the OKC is small and its image is similar to the jawbone SBC image, it cannot be distinguished by its size alone.

Among the texture features, GLCM contrast, NGTDM contrast, and GLCM variance had the highest correlation coefficients (Rs). GLCM is a matrix whose number of rows and columns represents the gray value, and the cell contains the number of times the gray value has a certain relationship (angle, distance, etc.). It is also known as a second-order histogram, and describes the two-by-two arrangement of voxels. GLCM not only reflects the distribution of brightness, but also reflects the distribution of position between pixels with the same or near-same brightness.²² Contrast refers to a measure of the local intensity change. A large contrast indicates that the intensity values between adjacent voxels have a large difference and that the visual effect is apparent. Variance reflects the measurement of pixel values and mean deviation. When the grayscale in the image changes greatly, the variance and standard deviation are large. In OKCs, the contents of the cyst cavity have a small difference in grayscale, and the change is relatively uniform. In contrast, SBCs contain blood or air, bone crest, bone fragments, and so on, and thus, the grayscale changes greatly because of which the texture features such as GLCM contract and GLCM variance are high.

NGTDM describes the difference between each voxel and the adjacent voxel.²³ Contrast reflects the change of spatial brightness. High contrast indicates that the image has a larger grayscale range and that there is more change in gray values around a certain point. The SBC sac cavity may have air density, blood or fluid density, and the density of bone crests or scattered bone fragments. These three types of densities have large differences, and so does the gray value changes. In OKCs, the gray values of keratin, sac fluid, and the sac wall are largely different. This observation explains that there is more gray value change around a certain point in SBC; therefore, the value of contrast is higher than that of OKC.

In order to obtain more comprehensive and accurate experimental results, we generated ROI along the boundary of the lesion, so that as many features as possible can be obtained. When the manual delineation was on the inter side-of the cyst, part of the information of cyst wall and shape would be lost, making the result inaccurate. On the other hand, excess bone tissue information would be brought into the margin if we outlined on the intra side, which may also interfere with the results. Precisely, the radiomics features of the boundary are important for they reflecting the difference between OKC and SBC, due to their distinctly histological characteristics. Therefore, we segmented just along the boundary of the cystic lesion in this study. To take a more comprehensive consideration of all the features of the image under various parameters, the four commonly parameters R, Scale, Algo and Ng were also used in extracting texture features in present study.

This study aimed to screen the radiomic features of simple bone cysts of the jaws. These features may help us explore new strategies for diagnosis. However, several limitations of the study have to be addressed. First, this was a single-center retrospective study. Second, SBC is a rare non-odontogenic cyst. Therefore, a small sample data set was applied in the study. A prospective randomized study with a larger data set to improve the proposed prediction model is needed in further studies.

In conclusion, this study primarily showed the potential ability of radiomics to differentiate SBCs and other odontogenic cysts. The texture features of radiomics quantitatively described the differences between SBC and OKC, and the three texture features (GLCM contrast, NGTDM contrast, and GLCM variance) showed the highest correlation coefficients, which were candidate radiomic features for diagnosing SBC.