Texture analysis of diffusion kurtosis imaging for differentiating malignant from benign sinonasal lesions: added value to conventional imaging features

Abstract

Objectives:

To evaluate the performance of texture analysis (TA) of diffusion kurtosis imaging (DKI) in differentiating malignant from benign sinonasal lesions, and its added value to the conventional imaging features.

Methods:

Fifty-eight patients with malignant and 40 patients with benign sinonasal lesions were retrospectively enrolled. Conventional CT and MRI features were reviewed. Texture parameters were obtained and compared between two groups. Multivariate logistic regression analysis was used to identify the most valuable variables. Receiver operating characteristic curves were performed to assess the differentiating performance of independent variables and their combination.

Results:

There were significant differences in tumor necrosis, bone erosion and soft tissue invasion between the two groups (all p < 0.05). There were significant differences in the 10th and entropy of Apparent diffusion coefficient map, the mean, 10th and entropy of D map, the mean and 90th of K map between the two groups (all p < 0.002). The bone erosion, entropy of D, and mean of K were independent variables associated with malignant tumors. Receiver operating characteristic analyses indicated that the combination of three features possessed better differentiating performance than bone erosion alone (p = 0.003).

Conclusion:

TA of DKI could supply incremental value to conventional imaging features for pre-operative differential diagnosis between benign and malignant sinonasal lesions.

Advances in knowledge:

The present study is the first to combine conventional imaging features and the TA of DKI in the differential diagnosis between benign and malignant sinonasal lesions. Our findings suggest that TA of DKI could supply incremental value to conventional imaging features.

Introduction

There is a broad spectrum of benign and malignant lesions in the nasal cavity and paranasal sinuses. ¹ Surgery or follow-up is the first-line treatment strategy for patients with benign sinonasal lesions, while comprehensive multimodal therapies are usually suggested for the patients with malignant sinonasal lesions. ² Therefore, accurate differential diagnosis between benign and malignant lesions is crucial for establishing individual treatment regimens. Differential diagnosis based on the clinical symptoms is difficult, because benign and malignant sinonasal lesions can lead to mimicking clinical presentation, and also be similar with that caused by chronic inflammation. ³ Thus, to find a simple and effective method for establishing an accurate differential diagnosis is urgently needed in clinical practice.

Endoscopic sinonasal resection biopsy is commonly used for pre-operative diagnosis, but its sensitivity is limited due to surrounding inflammatory tissue. ^4,5 Conventional CT and MRI are commonly used for the non-invasive assessment of sinonasal tumors. ⁶ However, the morphological features provided by CT and MRI are not enough to evaluate biological characteristics of sinonasal lesions. ⁷ Recently, diffusion-weighted imaging (DWI), which allows the characterization of water molecules diffusion in tumor tissues, has been proven to be helpful in distinguishing between benign and malignant lesions. ⁸ However, apparent diffusion coefficient (ADC) derived from the Gaussian diffusion model cannot accurately reflect the complicated water diffusivity in living tissues. ⁹ Previous studies have demonstrated that diffusion kurtosis imaging (DKI), as a non-Gaussian diffusion-weighted model, could better reflect the complexity and heterogeneity of tissue microstructure than DWI. ^9–11 Previous study reported that the mean values of DKI-derived parameters outperformed DWI and its derived ADC value in the differentiation between benign and malignant sinonasal lesions. ¹¹

Texture analysis (TA) which characterizes an image by analyzing the distribution and relationship of pixel gray levels, has attracted an increasing attention. ¹² It provides more objective and quantitative information regarding tumor heterogeneity beyond visual inspection. ¹² Several researches have reported promising results about applying TA of DKI functional maps for assessing the pathological features of breast lesions, prostate cancers, ovarian tumors and cervical cancer. ^13–16 However, to the best of our knowledge, few study has applied TA of DKI to differentiate malignant from benign sinonasal lesions until now.

Therefore, the purpose of our study is to evaluate the performance of TA of DKI in differentiating malignant from benign sinonasal lesions, and its added value to the conventional imaging features.

Methods and materials

Patients

This study was approved by our institutional review board, and informed consents were waived due to the retrospective nature. From January 2018 to December 2021, 121 patients with benign or malignant sinonasal lesions underwent head and neck MRI in our department. We included the patients who: (1) had no previous history of biopsy, surgery or any other treatment; (2) underwent conventional CT, MRI and DKI for pre-treatment evaluation; (3) had an adequate imaging quality without obvious artifacts; (4) had a final pathological diagnosis. Patients were excluded in cases of previous history of biopsy (n = 7), absence of CT scan (n = 5), absence of DKI sequences (n = 6), metal or motion artifacts that impaired the image quality (n = 3), and absence of a histological diagnosis (n = 2). Finally, 58 patients (male: female = 37: 21; mean age, 53.8 ± 15.6 years) with malignant sinonasal lesions, and 40 patients (male: female = 24: 16; mean age, 51.0 ± 17.4 years) with benign sinonasal lesions were enrolled in this study. Detailed pathological compositions in malignant group included lymphoma (n = 20), squamous cell carcinoma (n = 24), adenocarcinoma (n = 5), melanoma (n = 3), olfactory neuroblastoma (n = 3), adenoid cystic carcinoma (n = 2), and chondrosarcoma (n = 1). And benign group included: sinonasal papillomas (n = 21), inflammatory polyps (n = 11), hemangioma (n = 4), schwannoma (n = 2), respiratory epithelial adenomatoid hamartoma (n = 1), and plemorphic adenoma (n = 1).

CT imaging

All patients underwent CT scans using 16 or 64 channels (Siemens Healthineers, Germany) with or without contrast. Detailed scan parameters were as follows: 110 kV tube voltage; 250 mAs effective tube current; slice thickness, 1.5 mm. Coronal and sagittal CT images of both bone and standard soft tissue windows were routinely reformatted for further assessment.

MRI scan

MRI was performed using a 3 T scanner (Verio, Siemens Healthcare Sector, Erlangen, Germany) with a 12-channel head and neck coil. MRI protocol included the following sequences: (1) unenhanced axial T ₁ weighted imaging (repetition time [TR]/echo time [TE] = 811/7.1 ms, section thickness = 4 mm, field of view [FOV] = 220 × 220 mm, matrix = 384×384); (2) axial T ₂ weighted imaging with fat saturation (TR/TE = 4000/87 ms, section thickness = 4 mm, FOV = 220×220 mm, matrix = 384×384); (3) coronal T ₂ weighted imaging (TR/TE = 3800/88 ms, section thickness = 4 mm, FOV = 220×220 mm, matrix = 384×384), and (4) contrast-enhanced axial T ₁ weighted imaging (CE-T ₁WI) (TR/TE = 811/7.1 ms, section thickness = 4 mm, FOV = 220×220 mm, matrix = 384×384). For CE-T ₁WI, a standard dose of 0.1 mmol/kg of gadolinium-diethylene triamine pentaacetic acid (Omniscan, GE Healthcare, Dublin, Ireland) was administrated at a rate of 4 ml s⁻¹, followed by a 20 ml normal saline.

Readout-segmented echo planar imaging sequence (RESOLVE) was used for DKI scan. Detailed imaging parameters were showed as follows: diffusion schema, Stejskal-Tanner; fat suppression, frequency selective; b values, 0, 500, 1000 and 1500 s/mm²; orthogonal directions, 3; TR/TE, 5060/76 ms; slice number, 20; number of excitations, 1; FOV, 220 × 220 mm; slice thickness, 4 mm without gap; matrix, 224 × 224; phase-encoding direction, anteroposterior; echo spacing, 0.4 ms; number of readout segments, 5. Total acquisition time was 4 min 45 s.

Image analyses

Conventional CT and MRI features were assessed by two radiologists (with 8 and 4 years of experience in head and neck radiology, respectively) who were blinded to the study design and pathological information. Following qualitative image features, including location, shape, necrosis, bone erosion, soft tissue invasion, signal intensity on T ₁WI, uniformity of T ₁WI, signal intensity on T ₂WI, uniformity of T ₂WI, degree of enhancement, and uniformity of enhancement, were evaluated. Tumor location was defined as nasal cavity, nasal sinus or both. Tumor shape was classified as regular or irregular. Tumor border was noted as either clear or unclear. Compared with that of the adjacent muscle, the signal intensity of the solid part of mass on T ₁WI and T ₂WI was classified as hypointense, isointense, and hyperintense. Degree of enhancement was classified as slight (similar to muscles), moderate (greater than muscles) and marked (similar to the mucosa). ¹⁷ The uniformity of T ₁WI, T ₂WI and enhancement was used to evaluate the heterogeneity of sinonasal tumors.

Imaging data were post-processed offline with an in-house software (FireVoxel; CAI²R; New York University, NY) using monoexponential and DKI model. The monoexponential model is mathematically expressed as follows: S_b/S₀ = exp(-b*ADC). And DKI model is mathematically expressed as follows: S_b/S₀ = exp (-bD + $\frac{1}{6}$ b²D²K). D is apparent diffusion for Gaussian distribution, and K is apparent kurtosis coefficient (a dimension-less parameter). ¹¹ Imaging data of 2 b values (0 and 1000 s/m²) were processed using monoexponential model to measure ADC value while data of 4 b values (0, 500, 1000, and 1500 s/mm²) were processed using DKI model to measure D and K value. Regions of interest (ROIs) were also determined by the above-mentioned two radiologists. Based on diffusion images (b₁₀₀₀ map), ROIs were manually delineated around all the slices encompassing as much as tumor area. Three-dimensional (3D) volumes of interest (VOIs) were automatically constructed by summing ROIs drawn in each section. The VOIs would then be automatically copied to the corresponding parametric maps using the same software. Large necrotic, cystic, hemorrhagic areas and surrounding blood vessels were excluded referred to T ₂WI and contrast-enhanced T ₁WI in a manual approach (Figure 1). To reduce the influence of partial volume averaging, ROIs were slightly smaller in size than actual tumor sizes. After placing ROIs, the following texture parameters of ADC, D and K map were automatically obtained: mean, 10th, 90th, skewness, kurtosis, and entropy, which are based on histogram analysis and the gray-level co-occurrence matrix (GLCM) method. ¹⁸ Measurement results of the two radiologists were used to assess inter-reader reproducibility. The average value of the measurement results from the two radiologists was adopted into the statistical analysis.

Figure 1.

Schematic image (c) of region of interest placement on DK images (b = 1000 s/mm²), referred to T ₂WI (a) and contrast-enhanced T ₁WI (b). DK, diffusion kurtosis; T ₁WI, T ₁ weighted imaging; T ₂WI, T ₂ weighted imaging

Statistical analyses

Numeric data were averaged over all patients, and reported as mean ± standard deviation. Kolmogorov–Smirnov test was used for normally distributed analysis. Fisher’s exact test was applied to assess the differences of gender and each qualitative imaging feature between two groups. The differences in age between two groups were compared with an unpaired t-test. Texture parameters of DKI were compared between two groups using unpaired Student’s t-test. The significance threshold for difference was set at p < 0.0028 (0.05/18) after Bonferroni multiple comparison correction. Multivariate logistic regression analysis was used to identify the most valuable variables that were predictive of malignant sinonasal lesions. Based on the independent parameters, we established a combination model using the method of logistic regression. Receiver operating characteristic (ROC) curves were performed to assess the performance of those independent parameters and their combination in the differential diagnosis between benign and malignant sinonasal lesions. Area under the ROC curve (AUC), sensitivity, and specificity were calculated. Inter-reader agreement was evaluated using intraclass correlation coefficient (ICC) with 95% confidence intervals. ICC was classified as excellent (≥ 0.81), good (0.61–0.80), moderate (0.41–0.60), and poor (<0.40). All statistical analyses were performed by using MedCalc (v. 13.0; Mariakerke, Belgium) and SPSS (v. 25.0; IBM Corp., Armonk, NY). A two-sided p-value less than 0.05 was considered to be statistically significant.

Results

There were no significant differences in patient gender (p = 0.832) and age (p = 0.555) between benign and malignant sinonasal groups.

Table 1 summarizes the frequency distribution of conventional CT and MRI features between two groups. There were significant differences in tumor necrosis (p = 0.001), bone erosion (p < 0.000), and soft tissue invasion (p = 0.008). While no significant difference in tumor location, shape, signal intensity on T ₁WI, uniformity of T ₁WI, signal intensity on T ₂WI, uniformity of T ₂WI, degree of enhancement, and uniformity of enhancement was found (all p > 0.05).

Table 1.

CT and MRI features of malignant and benign sinonasal lesions

Parameters	Malignant	Benign	p
Location			0.390
Nasal cavity	14	16
Nasal sinus	12	10
Both	32	14
Shape			0.152
Irregular	52	31
Regular	6	9
Necrosis			0.001
Yes	45	17
No	13	23
Bony erosion			0.000
Yes	41	2
No	17	38
Soft tissue invasion			0.008
Yes	36	8
No	22	32
Signal intensity on T 1WI			0.823
Low	40	29
Iso	18	11
Uniformity of T 1WI			0.815
Yes	44	29
No	14	11
Signal intensity on T 2WI			0.052
Low	2	12
Iso	34	16
High	22	6
Uniformity of T 2WI			0.510
Yes	18	15
No	40	23
Degree of enhancement			0.408
Moderate	35	20
Marked	23	20
Uniformity of enhancement			0.825
Yes	17	13
No	41	27

T ₁WI, T ₁ weighted imaging; T ₂WI, T ₂ weighted imaging.

Excellent inter-reader agreements were achieved in the measurement of TA parameters (ICC, 0. 819–0.868). Detailed TA parameters of DKI are summarized in Table 2. As to TA parameters from ADC map, there were significant differences in the 10th and entropy between two groups (all p < 0.002). As to TA parameters from D map, there were significant differences in the mean, 10th and entropy between two groups (all p < 0.002). As to TA parameters from K map, the mean and 90th were higher in the malignant group than that in the benign group (all p < 0.002).

Table 2.

Comparison of texture parameters between malignant and benign groups

Parameters		Malignant	Benign	p
ADC	Mean	1.128 ± 0.445	1.334 ± 0.459	0.003
	10th	0.672 ± 0.295	0.899 ± 0.363	0.001
	90th	1.677 ± 0.498	1.928 ± 0.533	0.008
	Skewness	0.891 ± 0.722	0.652 ± 0.783	0.012
	Kurtosis	0.925 ± 0.937	0.826 ± 1.235	0.039
	Entropy	3.933 ± 0.354	3.842 ± 0.266	<0.001
D	Mean	1.354 ± 0.329	1.839 ± 0.456	<0.001
	10th	0.861 ± 0.256	1.275 ± 0.440	<0.001
	90th	1.926 ± 0.528	2.407 ± 0.558	0.004
	Skewness	0.994 ± 0.969	0.508 ± 0.878	0.080
	Kurtosis	1.586 ± 1.755	0.472 ± 1.124	0.010
	Entropy	3.989 ± 0.277	3.880 ± 0.268	<0.001
K	Mean	0.856 ± 0.190	0.612 ± 0.159	<0.001
	10th	0.220 ± 0.215	0.148 ± 0.146	0.173
	90th	1.409 ± 0.334	0.976 ± 0.285	<0.001
	Skewness	−0.156 ± 0.393	−0.009 ± 0.767	0.439
	Kurtosis	−0.299 ± 1.926	0.809 ± 1.783	0.047
	Entropy	3.925 ± 0.246	3.898 ± 0.241	0.703

ADC, apparent diffusion coefficient.

Multivariate logistic regression analysis showed that bone erosion, the entropy of D and the mean of K were independent differentiating variables (p = 0.032, 0.019 and 0.020, odds ratio = 1.368, 10.009 and 2.015, respectively). Representative cases of benign and malignant sinonasal lesions are shown in Figure 2. ROC analysis indicated that the combination of those three features had better diagnostic performance (AUC, 0.948; sensitivity, 96.6%; specificity, 80%) than bone erosion alone (AUC, 0.778; sensitivity, 65.5%; specificity, 90%) (p = 0.003). (Figure 3).

Figure 2.

(a–e) A 43-year-old male with left sinonasal papillomas. Axial CT (a), T ₂WI (b) and CE-T ₁WI (c) showed a mass located in left maxillary sinus and nasal cavity without bone erosion. The entropy of D (d) was 4.014, and the mean of K (e) was 0.700. (f–j) A 58-year-old male with right sinonasal NHL. Axial CT (f), T ₂WI (g) and CE-T ₁WI (h) showed an infiltrative mass located in right ethmoidal sinus and orbit with bone erosion. The entropy of D (i) was 4.858, and the mean of K (j) was 0.768. CE-T ₁WI, contrast-enhanced T ₁ weighted imaging.

Figure 3.

Diagnostic performance of bone erosion alone and the combination model.

Discussion

Correctly differentiating malignant from benign sinonasal lesions are essential in the individual treatment strategy. ² Diagnostic performance of conventional CT and MRI features have been reported in previous studies. ^6,7 However, qualitative assessment of imaging features is a subjective process, and the underlying information beyond the imaging is not fully excavated. As a promising approach, TA can provide more objective and quantitative information regarding tumor heterogeneity beyond visual inspection. ¹² Therefore, in present study, we aimed to evaluate the performance of the combination of conventional imaging features and the TA parameters of DKI functional maps in the differential diagnosis between benign and malignant sinonasal lesions. As a result, we found that the combination of bone erosion, the entropy of D, and the mean of K might be the most promising parameter, and TA of DKI could provide added value to conventional MRI features in the differential diagnosis between malignant and benign sinonasal lesions.

Our study found that tumor necrosis, bone erosion and soft tissue invasion were the significant qualitative imaging features assocaited with the malignant sinonasal lesions. The differential diagnostic accuracy is relatively low, because it depends on the radiologists’ experience and ability to interpret the CT and MR images. Previous studies also showed that bone involvement, tumor necrosis and involvement of the surrounding structures indicated a higher probability of malignancy than the average, with an accuracy of 64.5%. ^6,7,19 At present study, there was no significant difference in MRI signal intensity, uniformity and degree of enhancement between the malignant and benign sinonasal groups. Khaled et al also reported that MRI signal intensity and enhancement pattern in sinonasal tumors are non-specific, which was similar with our study. ¹⁹ These results indicate that the morphological features are still difficult to effectively diagnose malignant and benign sinonasal lesions.

DWI was commonly used for the discrimination between malignant and benign sinonasal lesions in clinical practice. ^8,11,19 In the present studies, the mean ADC value of benign sinonasal lesions was significantly higher than that of malignant lesions, however, the accuracy was not high. ^8,11,19 DKI and quantitative metrics derived from DKI have been proven to be useful for quantifying the degree of non-Gaussian distribution. ²⁰ D represents the corrected ADC accounting for non-Gaussian diffusion behavior, and K represents the apparent kurtosis coefficient which denotes the deviation of tissue diffusion from a Gaussian model. ²¹ As to DKI, Jiang et al reported that the accuracy of the mean K value (87.7%) was significantly better than that of the ADC (74%) and D value (80.2%). ¹¹ However, mean value ignored the complementary information on the complexity and heterogeneity in the tissue, which are imperceptible to the human naked eye. ¹⁶ In agreement with previous studies, we found that the mean ADC and D value was significant lower in malignant sinonasal lesions than that in benign mimics. This may be attributed to the enlarged and hyperchromatic nuclei, heteromorphism of nuclear contour, and hypercellularity of the malignant sinonasal tumors. ¹¹ And the lower 10th of ADC and D value was found in malignant sinonasal lesions in this study, the reason may be that the low-percentile ADC and D values represented the focal tumor area with high cell density, which usually represented the most aggressive component within the mass. ²⁰ Our study also showed that the entropy of D map was the independent differentiating parameter, and malignant sinonasal lesions showed a statistically higher entropy of D map than the benign group. Entropy was a crucial texture parameter associated with the intralesion heterogeneity. It has been proved to be a useful index for differentiating between a lot of types of malignant and benign tumors. ²²

As to K map, the mean and 90th were higher in the malignant group than that in the benign group, and mean K value was the independent differentiating parameter. K was proportional to the heterogeneity and complexity of microstructure of tumors. ¹¹ Therefore, malignant masses would show higher K than benign mimics. Our study results were consistent with previous findings in TA applied on DKI to assess pathological features in breast lesions, prostate cancers, ovarian tumors and cervical cancer. ^13–16 A combination of bone erosion, the entropy of D, and the mean of K for the diagnosis of the malignant sinonasal lesions yielded an optimal performance. This result indicated that the combination of conventional images and the TA parameters of DKI functional maps could help to establish the accurate diagnosis between benign and malignant sinonasal lesions.

Moreover, RESOLVE technique was used for DWI scan in our study. Compared with the widely used single-shot echo-planar (SS-EPI) technique, RESOLVE showed superiority in reducing distortion and artifact, and improving the overall imaging quality. High image quality might be an important cornerstone for subsequent image analysis in sinonasal lesions. ²³

This study has several limitations should be noted. Firstly, the sample size was relatively small. The wide heterogeneity of the included sinonasal tumor subtypes differed in their intrinsic morphologic features and DKI parameters. The heterogeneity might affect the diagnostic performance of the qualitative assessment of T1, T2 signal characteristics and enhancement. A higher number of the included sinonasal tumor subtypes was needed in future study. Secondly, we did not correlate the TA parameters of studied sinonasal tumors with the corresponding pathologic evaluation. Thirdly, other DKI parameters like AD, AK and FA, and other functional MRI techniques were not combined used for pre-treatment assessment. Lastly, further radiomics analyze were not performed in this study due to limited sample size. Future study integrating more study population, more functional MRI modalities, and radiomics analysis would be more attractive.

In conclusion, our preliminary study indicated that TA of DKI functional maps could supply added value to conventional imaging features in the differential diagnosis between benign and malignant sinonasal lesions. It might serve as an important reference in establishing individualized treatment plan for the patients with sinonasal lesions.