Enhanced CT-based texture analysis and radiomics score for differentiation of pleomorphic adenoma, basal cell adenoma, and Warthin tumor of the parotid gland

Abstract

Objective:

To evaluate the diagnostic performance of computed tomography (CT) radiomics analysis for differentiating pleomorphic adenoma (PA), Warthin tumor (WT), and basal cell adenoma (BCA).

Methods:

A total of 189 patients with PA (n = 112), WT (n = 53) and BCA (n = 24) were divided into a training set (n = 133) and a test set (n = 56). The radiomics features were extracted from plain CT and contrast-enhanced CT images. After dimensionality reduction, plain CT, multiphase-enhanced CT, integrated radiomics signature models and radiomics score (Rad-score) were established and calculated. The receiver operating characteristic (ROC) curve analysis was taken for the assessment of the model performance, and then comparison was conducted among these models. Decision curve analysis (DCA) was adopted to assess the clinical benefits of the models. Diagnostic performances including sensitivity, specificity, and accuracy of the radiologists were evaluated.

Results:

Seven, nine, fourteen, and fourteen optimal features were used to constructed plain scan, arterial phase, venous phase, and integrated radiomics signature models, respectively. ROC analysis showed these four models were able to differentiate PA from BCA and WT, with the area under the ROC curve (AUC) values of 0.79, 0.90, 0.87, and 0.94 in the training set, and 0.79, 0.89, 0.86, and 0.94 in the test set, respectively. The integrated model had better diagnostic performance than single-phase radiomics model, but it had similar diagnostic performance to that of the radiomics model based on the arterial phase (p > 0.05). The sensitivity, specificity, and accuracy in the diagnosis of PA were 0.86, 0.46, and 0.70 for the non-subspecialized radiologist and 0.88, 0.77, and 0.84 for the subspecialized radiologist, respectively. Six venous phase parameters were finally selected in differentiating WT from BCA. The predictive effect of the model was favorable, with AUC value of 0.95, sensitivity of 0.96, specificity of 0.83, and accuracy of 0.92. The sensitivity, specificity, and accuracy in the diagnosis between WT and BCA were 0.26, 0.87, and 0.45 for the non-subspecialized radiologist and 0.85, 0.58, and 0.77 for the subspecialized radiologist, respectively.

Conclusion:

The CT-based radiomics analysis showed favorable predictive performance for differentiating PA, WT, and BCA, thus may be helpful in the clinical decision-making.

Introduction

Parotid gland (PG) tumors are not common neoplasms, accounting for about 2 to 3% of all head and neck tumors, among which some 75–80% are benign. ^1,2 Benign parotid tumors mainly originate from epithelial and non-epithelial tissues. The most common benign tumors of epithelial origin are pleomorphic adenoma (PA), Warthin tumor (WT) and basal cell adenoma (BCA), which usually require different treatment because of different biological characteristics. For instance, approximately 5 to 10% of PAs carry a risk of malignant transformation if there is any delay for the excision and can recur after surgery. ^3,4 Thus, as for those patients with PA, partial superficial parotidectomy is recommended. The malignant transformation of WT rarely occurs, less than 1% of cases, and in the clinical practice, enucleation or conservative observation is recommended. ⁵ BCA can be classified into four histologic subtypes according to cellular growth patterns, including solid (most frequent form), tubular, trabecular, and membranous. The membranous type of BCA may have a comparatively high tendency to recur, while malignant transformation can occur in all four types. For the primary treatment of BCA in the cases where parotid affectation exists, surgical excision via a suprafacial or total parotidectomy is conducted. ⁶ Accurate preoperative diagnosis of PA, WT, and BCA would be significant in the consideration of an optimal individualized surgical planning and provides useful information for prognosis counseling.

Imaging tools, including magnetic resonance imaging (MRI), and computed tomography (CT), are commonly applied to examine the parotid region. MRI provides extremely good soft tissue resolution with extra information on the cellularity, organization, vascularity, neoangiogenesis, and microbleeds. Recently, there has been an increased interest in using the advanced MRI techniques including dynamic contrast-enhanced (DCE) MRI, diffusion-weighted imaging (DWI), and susceptibility-weighted imaging (SWI), to improve the distinction of parotid tumors. ^7,8 Contrast-enhanced CT may increase the incidence of contrast media-related adverse reactions, but it can also help clinicians to distinguish between benign and malignant parotid tumors and assess tumor boundaries, especially in the deep lobe. Different enhancement patterns of PA and WT can help improve the diagnostic efficiency, particularly gradual enhancement for PA and washout pattern enhancement for WT. ⁹ A previous study showed that energy spectrum CT had better discrimination ability compared to conventional CT in differentiating benign PG tumors. ¹⁰ However, parotid gland benign tumors often exhibit overlapping imaging characteristics, in addition, the conclusions mainly depend on the subjective judgment of radiologists, which is of low efficiency and brings difficultly for the radiologists with low levels of experience. Radiomics, as a medical tool can extract quantitative features from medical images via a variety of computational methods. It can reflect tumor heterogeneity by way of extracting, analyzing, and interpreting the quantitative imaging parameters and provide useful information for tumor classification, treatment efficacy, and evaluation of prognosis. ^11,12 Recently, several studies have reported successful applications of radiomics in head and neck tumors, such as the classification of primary oropharyngeal squamous cell carcinomas in accordance with human papillomavirus status based on CT texture analysis. ^13,14 In head and neck cancer radiotherapy, radiomics can predict the risk of severe toxicities, especially xerostomia, in addition to the response to radiochemotherapy or induction chemotherapy. ^15,16 Furthermore, it has been reported in several studies that the applications of radiomics analysis might be helpful in differentiating benign from malignant parotid tumors and PA from WT. ^17–19 A recent study ²⁰ used the CT-based radiomics nomogram for differentiating the lympho-associated benign and malignant PG tumors and demonstrated favorable discriminative ability.

However, there is a paucity of literature on the use of radiomics for precise classification of parotid PA, WT, and BCA. General consensus suggests that conventional multiphase contrast-enhanced CT imaging takes a crucial part in differentiating these three classes of benign PG tumors. In this study, we used conventional plain CT and multiphase-enhanced images to extract radiomics features and build a classification model for differentiation of PA, WT, and BCA. The aim of this study was to investigate the diagnostic performance of the prediction models based on plain CT and multiphase-enhanced CT, and to investigate whether the diagnostic performance could be improved by the integrated model in comparison with single-phase model and the radiologists.

Methods and materials

Patients

Study and review on the collected data for 357 patients who underwent CT for presurgical assessment of PG tumors from January 1, 2016, to December 30, 2020. One hundred eighty-nine patients diagnosed with PA (n = 112, 48 men and 64 women; mean age, 44 ± 14.8 years), WT (n = 53, 51 men and 2 women; mean age, 58.5 ± 12.1 years), and BCA (n = 24, 9 men and 15 women; mean age, 56.9 ± 9.7 years) were recorded in this study. The inclusion criteria were: (1) patients with pathologically confirmed PA, WT or BCA; (2) patients with complete clinic data; (3) patients who underwent non-contrast CT and contrast-enhanced CT of the head and neck within 7 days prior to the operation. The exclusion criteria were: (1) patients with a maximum tumor diameter < 5 mm; (2) images were not eligible for analysis because obvious artifacts, such as metallic or motion artifacts; (3) fine needle aspiration was carried out prior to imaging of the parotid tumor.

We combined BCA cohort with WT cohort, and performed radiomics analysis with PA cohort. All patients were randomly assigned to the training set and the test set in a ratio of 7:3, thereby 133 patients (79 PAs, 16 BCAs, 38 WTs) were allocated to the training set and 56 patients (33 PAs, 8 BCAs, 15 WTs) to the test set. Then we separately analyzed the texture characteristics of WT and BCA.

The approval for this retrospective study was obtained from the institutional review board of our hospital, and informed consent from the patients was not required in this study. The clinical and image data of patients were acquired from the hospital’s medical record system and the picture archiving and communication systems (PACS).

CT image acquisition

CT examinations were conducted on the 64-slice CT scanners (Discovery 750, GE Healthcare; optima 680, GE Healthcare) with the following parameters: 120 kVp tube voltage, 180–300 mAs tube current, matrix of 512 × 512, slice thickness of 2.5 mm, and section interval of 2.5 mm. The scanning ranged from zygomatic arch to mandible. The contrast-enhanced CT scan was made after intravenous injection of 70–80 ml of nonionic iodinated iodophosphol with 320 mg iodine per ml at 2.5 ml s⁻¹. Arterial and venous phase images were acquired at 25 sec and 70 sec after contrast agent injection.

Clinical and CT feature evaluation

Patients’ clinical parameters, including gender and age, were derived from admission records. Two radiologists (radiologist 1 with 8 years of experience, and radiologist 2 with 12 years of experience) who were unknown to the clinical data, including tumor location, shape, maximum diameter, distribution, calcification, cystic lesion, arterial enhancement, and enhancement pattern, independently reviewed and assessed CT feature analysis. For quantitative analysis of the enhancing characteristics, CT attenuation values (in HU) on arterial (25 s) and venous (70 s) CT scans were measured by placing the largest possible circular region of interest (3–120 mm²) within the solid portion of the lesion, but taking care to avoid the cystic area. The following diagnostic parameters in each tumor were calculated: A₀ was attenuation value on plain CT scan, A₁ was attenuation value on arterial-phase CT scan, and A₂ was attenuation value on venous-phase CT scan. The degree of arterial enhancement (A_d) as A_d = A₁ - A₀. We defined severe enhancement as A_d > 40 Hu, moderate enhancement as 20 Hu < A_d< 40 Hu, mild enhancement as 10 Hu < A_d < 20 Hu. Then the following diagnostic parameters were calculated: washout attenuation (A_w) as A_w = A₁ - A₂; and relative percentage enhancement washout ratio (RPEWR) as RPEWR (%) = (A_W / A₁) × 100. We defined the washout enhancement pattern as A₁ > A₂, and RPEWR > 10%; the platform type as A₁ > A₂, and RPEWR < 10%. When A₁ < A₂, we defined it as the gradual enhancement pattern.

Two other independent radiologists (radiologist 3: with 8 years of experience, but not subspecialized in head and neck radiology; radiologist 4: with 7 years of experience in head and neck radiology), unknown to the purpose of the study and the pathological outcomes, were requested to give their diagnostic impression among PA, WT or BCA.

Image segmentation and radiomics feature extraction

Two independent radiologists (radiologist 5 and radiologist 6 with 8 years of experience, trained in head and neck radiology) who were unknown to the clinical information and pathological outcomes selected the phase with the clearest lesion boundary after negotiation and then performed three-dimensional (3-D) segmentation of tumor region of interest (ROI) via ITK-SNAP software. All ROIs were then copied to the other phase images. The position, size, and shape of the tumor were manually adjusted. The contours of tumor were drawn slice-by-slice on axial CT slices meanwhile avoiding adjacent blood vessels and normal bone tissue. The imported image format was DICOM; and the exported format was NII. The image window width and window level were normalized before drawing the ROI (the window width was set to 350 Hu, and the window level to 40 Hu). An example of a manual segmentation was shown in Figure 1. Z-score normalization was conducted as preprocessing step for data to reduce the influence of different devices on the stability of imaging features. All ROIs from the plain CT, arterial, and venous phases images were uploaded into AK analysis software (Analysis Kit, V3.2.0.R, GE Healthcare) for feature extraction.

Figure 1.

Manual three-dimensional segmentation of a basal cell adenoma in a 65-year-old female. (a) segmentation on the axial slice. (b) three-dimensional volumetric reconstruction. (c) Case 1: pleomorphic adenoma of the left parotid gland in a 66-year-old female (white arrow). The mass shows mild contrast enhancement at the arterial phase, with CT value of 27–44 HU. (d) Case 2: Warthin tumor of the right parotid gland in a 68-year-old male (white arrow). The mass shows marked contrast enhancement at the arterial phase, with CT value of 90–110 HU. (e) Case 3: basal cell adenoma of the right parotid gland in a 65-year-old female (white arrow). The mass shows marked contrast enhancement at the arterial phase, with CT value of 90–140 HU.

Radiomics feature selection and development of the radiomics signature

Firstly, we evaluated the performance of CT radiomics analysis for differentiating PA from WT and BCA. We used three feature selection methods to eliminate the redundant and irrelevant features, i.e., Mann-Whitney U test, minimal redundancy maximal relevance (mRMR), and least absolute shrinkage and selection operator (LASSO). We used 10-fold cross-validation to tune the regularization parameter (λ) with minimum criteria in the training set. The average of the 10 CVs was used to evaluate if the model was overfitting, and the model of the one with the lowest validation loss was kept for the final model. Eventually, the plain CT, multiphase-enhanced CT, and integrated radiomics signature models were established. A radiomics score (Rad-score) was calculated for each patient. The Rad-score values of PA and the integrated cohorts in the training and test sets were compared. In the texture analysis of BCA and WT, the following four feature selection methods were applied, i.e., Mann-Whiney U-test, univariate logistic regression, mRMR, and multivariable logistic regression. After filtering the independently discriminative features, we constructed the final model and calculated the Rad-score.

Thirty randomly chosen CT images (10 PAs, 10 BCAs, and 10 WTs) were analyzed to assess the intraobserver and interobserver repeatability and reproducibility of radiomics features extraction. Radiologist 5 and radiologist 6 independently performed ROI segmentation on CT images during the same period in order to evaluate the features’ reproducibility between observers. To assess the repeatability within the observer, Radiologist 5 performed the same procedure two weeks later. The intraobserver and interobserver agreements were assessed using the intraclass correlation coefficient (ICC), with an ICC > 0.75 indicating good agreement. If there was a strong agreement, radiologist 5 performed image segmentation on the remaining samples.

Statistical analysis

R statistical software (version 3.5.1; https://www.r-project.org) was adopted for the radiomics statistical analysis. Wilcoxon statistical method was used to compare the Rad-score values of PA and integrated cohorts in the training and test sets. The receiver operating characteristic (ROC) curve was applied to evaluate the diagnostic performance of the models and to calculate the models’ specificity, sensitivity, accuracy, positive predictive value, and negative predictive value based on the Youden index. The Delong test was applied to compare the differences between the various area under the curve (AUC) values. Decision curve analysis (DCA) plots were generated with the “risk model decision analysis (rmda[R])” package.

SPSS version 23.0 was applied to analyze the clinical data and imaging features of all patients. Normally distributed quantitative data were analyzed by t-test whereas for non-normally distributed quantitative data the Mann-Whitney test was used. Chi-square or Fisher’s exact test was performed for qualitative data. A two-tailed p-value < 0.05 was considered to be statistically significant.

Results

Clinical and radiological features

In the PA and integrated cohorts, all data were divided into a training set and a test set. The clinical and radiological characteristics of the patients in the training and test sets were shown in Table 1. There was a statistically significant difference only in patients’ gender between the two groups. And the patients’ age, gender, tumor distribution, arterial enhancement, and enhancement pattern were significantly different between PA and integrated cohorts in both training and test sets (Table 2). Moreover, there were significant differences in the gender, tumor distribution, cystic degeneration, and enhancement pattern between the BCA and WT cohorts (Table 3).

Table 1.

Clinical and radiological characteristics of patients in the training and test sets

Variable	Training set (n = 133）	Test set (n = 56）	P-value
Age (year)	48.5 ± 15.1	52.9 ± 14.9	0.069
Gender (male/female)	69/64	39/17	0.024
Tumor Distribution (single/multiple)	120/13	52/4	0.564
Location (shallow/deep/across)	90/0/43	44/0/12	0.187
Maximum diameter (mm)	22.8 ± 7.5	23 ± 8.5	0.907
Calcification (yes/no)	2/131	1/55	1.000
Shape (round/non-rounded)	127/6	54/2	1.000
Cystic degeneration (yes/no)	65/68	27/29	0.934
Arterial Enhancement (mild/moderate/severe)	59/28/46	24/10/22	0.794
Enhancement curve (inflow/outflow/platform)	96/34/3	36/19/1	0.467

Table 2.

Clinical and radiological characteristics of pleomorphic adenoma (PA), basal cell adenoma (BCA), and Warthin tumor (WT) in the training and test sets

Variable	Training set (n = 133)			Test set (n = 56)
	PA	BCA&WT	P-value	PA	BCA&WT	P-value
Age (year)	42.6 ± 15	56.7 ± 10.9	＜0.001	47.3 ± 14.2	61.0 ± 12.1	＜0.001
Gender (male/female)	29/50	40/14	＜0.001	19/14	20/3	0.019
Distribution (single/multiple)	79/0	41/13	＜0.001	33/0	19/4	0.024
Location (shallow/deep/across)	51/0/28	39/0/15	0.353	25/0/8	19/0/4	0.743
Maximum diameter (mm)	22.6 ± 7.4	23.3 ± 7.7	0.811	21.3 ± 8.1	25.5 ± 8.5	0.053
Calcification (yes/no)	1/78	1/53	1.000	0/33	1/22	0.411
Shape (round/non-rounded)	74/5	53/1	0.400	31/2	23/0	0.507
Cystic degeneration (yes/no)	34/45	31/23	0.104	13/20	14/9	0.114
Arterial Enhancement (mild/moderate/severe)	48/20/11	11/8/35	＜0.001	21/5/7	3/5/15	0.001
Enhancement pattern (inflow/outflow/platform)	76/3/0	20/31/3	＜0.001	30/3/0	6/16/1	＜0.001

Table 3.

Clinical and radiological characteristics of basal cell adenoma (BCA) and Warthin tumor (WT)

Variable	WT (n = 53)	BCA (n = 24)	P-value
Age (year)	58.5 ± 12.1	56.9 ± 9.7	0.562
Gender (male/female)	51/2	9/15	＜0.001
Distribution (single/multiple)	36/17	24/0	0.002
Location (shallow/deep/across)	41/0/12	17/0/7	0.538
Maximum diameter (mm)	24.4 ± 7.9	22.8 ± 8.1	0.254
Calcification (yes/no)	1/52	1/23	0.529
Shape (round/non-rounded)	52/1	24/0	1.000
Cystic degeneration (yes/no)	36/17	9/15	0.012
Arterial Enhancement (mild/moderate/severe)	10/10/33	4/3/17	0.824
Enhancement curve (inflow/outflow/platform)	12/41/0	14/6/4	＜0.001

PA vs. BCA and WT

A total of 718 features showed good inter- and intraobserver agreement, with ICCs > 0.75. Radiologist five performed the feature extraction on the remaining samples because of strong agreement. First-, second-, and higher order texture features were extracted, including shape, gray histogram features, gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level dependence matrix (GLDM), gray-level size zone matrix (GLSZM), Gauss transform and wavelet transform. After feature dimensionality reduction, seven, nine, fourteen, and fourteen optimal features were used to construct plain scan, arterial phase, venous phase and integrated radiomics signature models, respectively (Figure 2). Rad-score was then calculated by these selected features weighted by their coefficients, and the Rad-scores were quite different between PA and integrated cohorts in both training and test sets (Supplementary Material 1). Moreover, the ROC curves of the four models demonstrated good predictive performance for both the training and test sets (Table 4 and Figure 3). ROC analysis showed that plain CT, arterial phase, venous phase, and integrated models were able to differentiate PA from BCA and WT, with the AUC values of 0.79, 0.90, 0.87 and 0.94 in the training set, and 0.79, 0.89, 0.86 and 0.94 in the test set, respectively. In the training set, Delong test showed that the AUC value of the integrated model was remarkably different from that of the plain CT and venous phase models (p < 0.001 and p = 0.036, respectively). In the test set, the AUC value of the integrated model was only remarkably different from that of the plain CT model (p = 0.04). The DCAs demonstrated that the integrated model had a higher overall net benefit than the other three models (Figure 4).

Figure 2.

Radiomics feature selection using the least absolute shrinkage and selection operator (LASSO) regression model. (a) plain scan model. (b) arterial phase model. (c) venous phase model. (d) integrated model.

Table 4.

Diagnostic efficacy of the four radiological models

Group/Model	Se	Sp	PPV	NPV	Accuracy	AUC	95% CI	P [a]
Training set
Plain CT	0.76	0.71	0.64	0.81	0.73	0.79	0.72–0.87	< 0.001
Arterial phase	0.93	0.73	0.70	0.93	0.81	0.90	0.85–0.95	0.114
Venous phase	0.80	0.81	0.74	0.85	0.80	0.87	0.81–0.93	0.036
Integrated group	0.94	0.77	0.74	0.95	0.84	0.94	0.91–0.98	-
Test set
Plain CT	0.70	0.70	0.62	0.77	0.70	0.79	0.67–0.92	0.040
Arterial phase	0.87	0.76	0.71	0.89	0.80	0.89	0.80–0.98	0.400
Venous phase	0.65	0.90	0.83	0.79	0.80	0.86	0.75–0.97	0.188
Integrated group	1.00	0.76	0.74	1.00	0.86	0.94	0.88–1.00	-

^aThe differences between AUC values in the integrated model were compared with that in the plain scan, arterial phase, and venous phase models by Delong’s test. Se, sensitivity; Sp, specificity; AUC, area under the curve; PPV, positive predictive value; NPV, negative predictive value; CI, confidence interval.

Figure 3.

The receiver operating characteristic (ROC) curves of the four models in the training (a) and test (b) sets, respectively.

Figure 4.

Decision curve analysis (DCA) for four models. The y-axis indicates the net benefit; the x-axis indicates threshold probability. The black line, blue line, yellow line, and green line represent the net benefit of the plain scan, artery phase, venous phase, and integrated models, respectively. The combined model had a higher overall net benefit in differentiating pleomorphic adenoma, basal cell adenoma, and Warthin tumor than the other three models.

The sensitivity, specificity and accuracy in the diagnosis of PA were 0.86, 0.46 and 0.70 for the non-subspecialized radiologist, and 0.88, 0.77 and 0.84 for the subspecialized radiologist, respectively.

BCA vs. WT

Texture analysis based on the plain CT and multiphase-enhanced CT showed that totally 2154 radiomics features were selected by Mann-Whitney U-test, and 415 features were retained by univariate logistic regression. Next, mRMR method was used to filter the redundancy features and obtain eight features (Figure 5). Finally, six most predictive features were retained through backward stepwise selection with the likelihood ratio test. Radiomics signature was eventually constructed by multivariable logistic regression. The predictive effect of the model was favorable, with an AUC value of 0.95, specificity of 0.83, sensitivity of 0.96, and accuracy of 0.92 (Figure 6).

Figure 5.

Identification of parotid basal cell adenoma and Warthin tumor by multiple parameters of CT texture analysis with ROC curves.

Figure 6.

The ROC curve of the radiomics signature in differentiating the basal cell adenoma and Warthin tumor.

The rad-score was calculated as follows: Rad-score = 2.21203943475944×(Intercept) + (4.89166638004698 × original_glcm_Id) + (1.35617355250434 × wavelet_HLL_gldm_Dependence Non-Uniformity Normalized) + (4.68288775352499 × original_glcm_Joint Energy) + (−3.30766515475012 × original_glrlm_Long Run Emphasis) + (3.53064532937874 × wavelet_LLL_glcm_Joint Entropy) + (−1.08608003217607 × log_sigma_3_0_mm_3D_firstorder_Skewness).

The sensitivity, specificity and accuracy in the diagnosis of WT were 0.26, 0.87 and 0.45 for the non-subspecialized radiologist, and 0.85, 0.58 and 0.77 for the subspecialized radiologist, respectively.

Discussion

Our results indicated that the multiphase combined model had a better diagnostic performance than single-phase radiomics models in differentiating PA from WT and BCA, but it had similar diagnostic performance to the arterial phase radiomics signature model. Moreover, our results also showed that texture analysis based on venous phase parameters might be helpful to distinguish WT from BCA.

Although there is a potential risk associated with CT radiation dose, multiphase contrast-enhanced CT is still used as the preoperative method for the evaluation of PG tumors. Because parotid gland lesions have a variable enhancement pattern. In this study, PA frequently showed mild to moderate enhancement in the arterial phase (94/112, 84%) and gradual enhancement pattern (106/112, 95%); the majority of WT showed rapidly significant enhancement (33/53, 62%) and a washout enhancement pattern (41/53, 77%). Previous studies have also shown that WT may present delayed enhancement, while PA a quick expurgation, ²¹ which was consistent with our data. In addition, BCA not only showed the characteristics of rapidly significant enhancement in arterial (17/24, 71%), but also showed both gradual, platform and washout enhancement patterns. ²² The mechanisms behind these features have not been fully understood. We presumed that the different histologic constituents, cellularity, extracellular space, vascular density, and leaky blood vessels might partly account for the various patterns of enhancement. Thus, diagnosis of these tumors may be challenging considering that PA, BCA, and WT present similar imaging features.

Radiomics, a recently technique, for characterizing intratumoral heterogeneity by extracting high-throughput features from medical imaging. ²³ Zhang et al ²⁴ reported that texture analysis based on arterial-phase images could be used for the preoperative, non-invasive diagnosis of PA and WT. In this study, the 3D whole tumor radiomics analysis based on both plain CT and multiphase-enhanced CT images was conducted to differentiate PA, WT, and BCA. The heterogeneous area of the tumor, including cystic, necrotic, and hemorrhagic areas, may help in the categorization and differentiation of tumors. One point should be noted was that the clinical features of the training and test sets were preferable to maintain balance. However, gender distribution between training and test sets was different in this study. All patients were randomly divided into the training and test sets in a ratio of 7:3 in our cohort. The randomization process may be resulted in an unequal sex distribution. But we suggested that it might be more important to control for consistency in the distribution of imaging features compared to gender distribution in this study. We selected 7, 9, and 14 effective features from plain CT, arterial phase, and venous phase by 10-fold cross-validation LASSO regression, respectively. At the same time, we also selected 14 optimal features in the integrated model, including three plain CT features (Mean Absolute Deviation, Skewness, GLCM_IMC1), eight arterial phase features (10th percentile, Skewness, 90th percentile, GLRLM_Short Run Emphasis, GLDM_Small Dependence Emphasis, GLCM_Inverse Variance, GLCM_Correlation), and three venous phase features (GLDM_Dependence Variance, GLCM_Inverse Variance, GLDM_Dependence Non-Uniformity Normalized). Mean absolute deviation refers to the mean distance of all intensity values from the mean value of the image array. Short run emphasis measures the distribution of short run lengths, with a greater value indicative of more fine textural textures. GLDM features also reflect heterogeneity and homogeneity, with a greater value of small dependence emphasis indicating less homogeneous textures, and a lower value of dependence non-uniformity normalized indicating more homogeneous textures. GLCM features that depict the spatial relationship between two neighboring pixels or voxels with predefined gray-level intensities, can show intratumor heterogeneity and may be associated with tumor malignancy. ¹⁹ Zhang et al ²⁴ reported that texture features, mainly comprised by GLCM, were able to distinguish benign from malignant PG tumors. Gabelloni et al ²⁵ reported that GLCM could not only distinguish between benign and malignant PG tumors, but also PAs from malignant tumors or WTs. Therefore, the intratumor heterogeneity may be used to explain the results of this study. PA contains varying histologic components of epithelial, myxoid, and fibrous elements, while WT consists of epithelia, monomorphic, oncocytic components, and lymphoid stroma. ²⁶ Moreover, BCA is a benign tumor comprised of a single layer of basal cell-like epithelial cells with a clear basement membrane-like structure, lacking the mesenchymal or cartilage-like matrix components which are often found in PA. ²⁷ WT has high vascularity and capillary-like vessel networks, while solid type BCA has numerous endothelial-line vascular channels with prominent small capillaries and venules. ²⁸ In addition, features including skewness, 10 percentile, and 90 percentile also showed good diagnostic performance. The differences of the 10th and 90th percentile also reflect the heterogeneity of tumor tissue composition or distribution. Skewness measures the asymmetry of the distribution of values about the Mean value. Higher skewness values have been related to tumor angiogenesis. ^29–31 WT has higher vascularity and microvessel permeability, while BCA has a rich capillary networks and venules. Therefore, the skewness of arterial and venous phases may have higher diagnostic efficiency.

Our study showed that the radiomics models based on plain CT and multiphase-enhanced CT texture features had favorable predictive value in distinguishing PA from WT and BCA, and the integrated model presented the best efficacy, with AUC value of 0.94. Compared with the other three models, the AUC values of the integrated models were only statistically significant with that of the plain CT models in both training and test sets, but not with that of the arterial phase models. This may be related to the obvious early enhancement of WT and BCA in the arterial phase, while mild enhancement of PA was observed. The characteristics of arterial hemodynamics had an important role in distinguishing PA from WT and BCA. The texture features were consistent with the routine findings. Moreover, when we compared our results to those in recent publications evaluating MR imaging for the distinguishment of PA and WT, our study showed better diagnostic performance. ¹⁷

In addition, texture analysis based on the venous phase parameters had a higher value in differentiating WT from BCA. The reason may be that some BCAs show continuous enhancement in the venous phase, while WTs often demonstrate a washout pattern of enhancement. Inverse difference (ID) based on the vein phase texture parameter showed the best diagnostic performance. It is a measure of the uniformity of the local grayscale distribution of an image. The more uniform gray levels would result in a higher ID value. Nevertheless, the lower of the ID value, the more irregular texture are.

The predictive radiomics models in our study revealed that the arterial phase model and the integrated model resulted in higher sensitivity for the diagnosis of PA in comparison with the diagnosis made by both non-subspecialized and subspecialized head and neck radiologists, and higher than 0.46 specificity and 0.70 accuracy by the non-subspecialized radiologist, but similar to 0.77 specificity and 0.84 accuracy by subspecialized radiologist. Vernuccio et al ³² reported that radiomics model based on combination of all MRI sequences had higher sensitivity to distinguish PA from WT in comparison with the diagnosis made by both non-subspecialized and subspecialized radiologists, but without a benefit from the aspect of specificity. There were some consistencies with our findings. Furthermore, the model based on venous phase parameters showed higher sensitivity and accuracy in differentiating WT from BCA in comparison with the diagnosis made by both non-subspecialized and subspecialized radiologists. Interestingly, the specificity of the model was higher than subspecialized radiologists, but similar to non-subspecialized radiologists. Therefore, if our results are validated in an external large-cohort study, we believe that the enhanced CT-based radiomics analysis can improve the differentiation of PA, WT, and BCA, and help the clinicians in decision making.

Radiomics analysis may be affected by several factors, including random variations in scanner and acquisition protocol, reconstruction settings, and ROI segmentation. ^33,34 If the data were from different manufacturers with different resolution and scan parameters, even different models of scanners from the same manufacturer, batch effect would exist and affect the value of the radiomics features, so data preprocessing such as z-score, combat and functional normalization and so on should be used to eliminate the batch effect first. However, these methods need further validation, because important textural information may be lost in this process. Manual segmentation has high interobserver variability, which can lead to the derivation of unstable radiomic features. Forde et al ³⁵ demonstrated that 7 and 10% of features were regarded to be unstable for the auto-segmentation and manual delineations respectively, with the larger sample size of 40 participants. The intraclass correlation coefficient (ICC) was used to evaluate intraobserver and interobserver agreement. In this study, we only observed interobserver and intraobserver variability between the two radiologists, with an ICC > 0.75 indicating good agreement. We did not use higher ICC cutoff values or 95% confidence intervals for verification of the reproducibility of these results. This was a limitation of the present study. The solution is to increase the number of observers and expand the sample size, or to use an accurate automatic method in order to obtain more robust features. Overfitting or underfitting is an important concern for radiomics. In this study, we selected optimal features using LASSO regression analysis, and used 10-fold cross-validation procedures, which have been shown to reduce the overfitting to a certain extent. Underfitting is a scenario in data science when a model is unable to classify data correctly and generate a high error rate in both the training and the validation sets. ³⁶ The consistency of AUC values on the training and test sets may indicate underfitting, and a good model should produce a small gap. In the present study, the training and test sets showed identical AUC values in the plain CT and integrated radiomics signature models, but with different confidence intervals. In addition, there were 7 and 14 optimal features in these two models, respectively, to reduce the underfitting caused by too few features, and the models had relatively high accuracy. Of course, such results remain to be further externally validated.

This study still has some limitations. First of all, it was a retrospective study that included a single institution and had a relatively small sample size, all of which may lead to bias. Secondly, the agreement analysis was conducted with about one-six of the study population, not all cases. Therefore, further multicenter studies with larger patient population are needed to validate the radiomics model. Thirdly, the present study used manual segmentation which was time consuming, and the results may be compromised due to interobserver variability. So, the generalizability of our results to other institutions and users is unclear. Fourthly, there was a lack of understanding on the potential relationship between specific texture features and histopathologic features. Finally, only three benign parotid tumors were analyzed, without expansion of analysis to other benign or malignant PG tumors. In order to increase the accuracy of diagnosis, it is planned to perform further studies to evaluate the radiomics analysis in other parotid gland lesions and try to apply other machine learning algorithms for data analysis in the near future.

Conclusion

In conclusion, our study demonstrated that CT-based radiomics analysis could be applied in differentiating PA, WT, and BCA. The integrated model had similar diagnostic performance to that of the arterial-phase radiomics model in differentiating PA from WT and BCA. Moreover, texture analysis based on the venous phase parameters had a higher value in differentiating WT from BCA. Therefore, the radiomics model may be taken as a non-invasive and valid tool to potentially complement conventional clinical imaging in clinical decision-making, although further validation is required before it can be widely used in clinical application.