Radiomics analysis for differentiating of cervical lymphadenopathy between cancer of unknown primary and malignant lymphoma on unenhanced computed tomography.

To investigate the usefulness of texture analysis to discriminate between cervical lymph node (LN) metastasis from cancer of unknown primary (CUP) and cervical LN involvement of malignant lymphoma (ML) on unenhanced computed tomography (CT). Cervical LN metastases in 17 patients with CUP and cervical LN involvement in 17 patients with ML were assessed by 18F-FDG PET/CT. The texture features were obtained in the total cross-sectional area (CSA) of the targeted LN, following the contour of the largest cervical LN on unenhanced CT. Values for the max standardized uptake value (SUVmax) and the mean SUV value (SUVmean), and 34 texture features were compared using a Mann-Whitney U test. The diagnostic accuracy and area under the curve (AUC) of the combination of the texture features were evaluated by support vector machine (SVM) with nested cross-validation. The SUVmax and SUVmean did not differ significantly between cervical LN metastases from CUP and cervical LN involvement from ML. However, significant differences of 9 texture features of the total CSA were observed (p = 0.001 - 0.05). The best AUC value of 0.851 for the texture feature of the total CSA were obtained from the correlation in the gray-level co-occurrence matrix features. SVM had the best AUC and diagnostic accuracy of 0.930 and 84.8%. Radiomics analysis appears to be useful for differentiating cervical LN metastasis from CUP and cervical LN involvement of ML on unenhanced CT.

INTRODUCTION

Enlarged cervical lymph nodes (LNs) have been reported to indicate malignancy in over 50% of cases.[1,2] Cervical LN metastasis from head and neck carcinoma and cervical LN involvement of malignant lymphoma (ML) in the head and neck are the most common disease associated with cervical LN malignancy.[2,3] Cancer of unknown primary (CUP) is defined as metastatic disease without evidence of a primary tumor on physical examination, endoscopy, or imaging. Cervical LN metastasis from CUP accounts for up to 7% of all head and neck carcinomas.[4,5] Discrimination between cervical LN metastasis from CUP and cervical LN involvement of ML is challenging when the primary cancer in the head and neck is not detected on initial evaluation. Fine-needle aspiration/biopsy are performed to confirm tissue characteristics of cervical lymphadenopathy in cases of suspected malignancy. Diagnosis of ML requires sufficient tissue for histological, immunophenotypic, and genetic studies to identify the ML subtype and determine the optimal treatment strategy.[6,7] In addition, patients with cervical LN enlargement might undergo random biopsy of the larynx and pharynx to identify the primary tumor site. Unfortunately, such invasive approaches are associated with a risk of complications and a possibility of insufficient materials.[6-9]

Computed tomography (CT) is performed to evaluate the presence of inflammation and metastasis, primary malignant tumor location, degree of extracapsular extension, and ML stage when cervical lymphadenopathy is identified clinically. The imaging findings for cervical LN metastasis from CUP and cervical LN involvement of ML on CT and magnetic resonance can be similar.[10] Therefore, differentiating between cervical LN metastasis from CUP and cervical LN involvement of ML is currently challenging. Recently, artificial intelligence has been applied to medical imagings.[11-17] Also, quantitative imaging analysis has been developed to improve diagnostic performance with increasing reproducibility and decreasing variability between radiologists. Texture analysis, a mathematical method used to calculate differences in gray-level patterns, has the potential to evaluate tumor characteristics, predict response to therapy, and determine prognosis in patients with malignant tumors.[18-25] However, no previous studies have described the differentiation between cervical LN metastasis from CUP and cervical LN involvement of ML using texture analysis. Therefore, we hypothesized that CT-based texture features using machine learning can discriminate between cervical LN metastasis from CUP and cervical LN involvement of ML. The aim of this study is to investigate the diagnostic accuracy of texture analysis in differentiating cervical LN metastasis from CUP and cervical LN involvement of ML on unenhanced CT.

MATERIALS AND METHODS

This retrospective study was approved by our institutional review board, which waived the need for informed consent from patients.

Subjects

Initially, 57 patients with histopathologically proven cervical LN metastasis from CUP (n=17) and ML (n=40) who underwent fluorine-18-2-fluoro-2-deoxy-D-glucose positron emission tomography (18F-FDG PET)/CT from April 2013 to April 2018 were identified.

Inclusion criteria for CUP with cervical LN metastasis

All patients with cervical LN metastases from CUP fulfilled the following inclusion criteria: 1) a physical examination and nasal endoscopy by a otolaryngologist and an 18F-FDG PET/CT before treatment were performed and failed to identify the original cancer site; 2) metastases of enlarged cervical LN(s) were proven histopathologically by neck dissection or fine-needle biopsy; and 3) the treatment for CUP was ultimately performed based on the diagnosis of cervical LN metastasis from CUP, and on a discussion by otolaryngologists and radiologists. Because some patients were treated prior to implementation of the recommendation to evaluate Epstein-Barr virus (EBV) and human papilloma virus (HPV) infection status for occult cervical cancer,[26] it was not necessary to evaluate EBV or HPV infection status for all patients.

Inclusion criteria for cervical LN involvement of ML

Seventeen patients were enrolled in this study based on the following criteria: 1) abnormal FDG uptake was visually identified in the enlarged cervical LN on 18F-FDG PET/CT and 2) ML was confirmed by fine-needle biopsy of the enlarged neck LNs.

FDG-PET/CT scanning

All patients who had fasted for at least 5 hours underwent whole-body 18F-FDG-PET/CT (the Biograph mCT-S(64)4R; Siemens Healthineers, Forchheim, Germany) from the vertex of the skull to the floor of the pelvis. 18F-FDG (3.7 MBq/kg BW, max 340MBq) was given intravenously. CT was performed 1 hour after 18F-FDG injection. The following scan parameters for the 64-row whole-body PET/CT scanner were used: tube voltage, 120 kVp; tube current, automatic exposure control (CARE Dose 4D); gantry rotation time, 0.5 sec; beam pitch, 1.5; imaging field of view, 500 × 500 mm; matrix, 512 × 512; and slice thickness, 2mm. All images were reconstructed with a B31f kernel. PET data were reconstructed using a three-dimensional (3D) iterative algorithm of ordered subsets expectation-maximization (OSEM) (2 iterations, 21 subsets).

18F-FDG PET image analysis

All PET data were reviewed on a commercially available workstation (Syngo via VB10; Siemens Healthineers). The boundary of volume of interest (VOI) for PET was semi-automatically drawn by the workstation in the largest enlarged cervical LN. The maximum standardized uptake value (SUVmax: SUV of the highest count within the VOI) and the mean SUV value (SUVmean: SUV of the mean count within the VOI) were calculated automatically by the workstation.

Image segmentation and texture analysis

On the axial planes of the CT images, the texture features were analyzed in the total cross-sectional area (CSA) of the target lesion. On each axial CT image, the regions of interest (ROIs) were placed to include the targeted LN. LNs were manually countered by consensus of two radiologists (*blinded* and *blinded*, who had 7 and 19 years of experience in radiology, respectively) using LifeX Software (https://www.lifexsoft.com, CEA, Saclay, France).[27] CT images with severe metal artifacts were not analyzed. Texture features were classified as follows: 4 histogram features, 6 gray-level co-occurrence matrix (GLCM) features, 11 gray-level run length matrix (GLRLM) features, 2 neighborhood gray-level different matrix (NGLDM) features, and 11 gray-level zone length matrix (GLZLM) features. A list of all texture features is provided in Appendix 1.

Statistical analysis

JMP 10.0.2 software (SAS Institute, Cary, NC, USA) was used for statistical analyses. Data were expressed as the mean ± standard deviation. A Mann-Whitney test was used to compare the SUVmax, SUVmean, and 34 texture features of the total CSA between the cervical LN metastases from CUP and the cervical LN involvement from ML. Receiver operating characteristic (ROC) curves and the area under the curve (AUC) of the ROC were calculated for each texture feature. A python-based support vector machine (SVM) with radial basis kernel in the machine learning library ‘scikit-learn’ (v0.16.1; http://scikit-learn.org) was implemented to evaluate accuracy and AUC for combinations of texture features with a nested cross-validation (10 repetition of 5-fold inner cross-validation and 10 repetition of 5-fold outer cross-validation) (Fig 1). A non-nested cross-validation has a bias that refers to the same data to optimize model parameters and evaluate model performances.[28] In a nested cross-validation, different data between tuning parameters and evaluating model performances are used to avoid overfitting: 1) an outer 5-fold cross-validation splits all data into 4 training sets and 1 test set; 2) an inner 5-fold cross-validation splits the training data into another 4 training set and 1 test set; 3) parameters are tuned in an inner 5-fold cross-validation that is repeated 10 times; 4) a remaining test set in an outer 5-fold cross-validation is used to evaluate SVM performance using the optimized parameters; and 5) the series of processes are performed in an outer 5-fold cross-validation, which is repeated 10 times.[29,30] The grid search method was used to select the optimized SVM parameters: C and gamma. The best combinations of the texture features were selected by a cursive feature elimination (RFE) method. RFE reduced the combination of texture features to specified number of top texture features according to importance.[31] All p-values less than 0.05 were considered to indicate significant difference.

Fig. 1

Machine learning using nested cross-validation

RESULTS

A total of 34 patients were included. Pretreatment characteristics of patients and tumors are shown in Table 1. Measurements of CT-related radiation dose were calculated by CT dose index (CTDI). The mean CTDIvol was 4.73±3.41 mGy.

Table 1

Characteristics of patients and tumors

	CUP		ML
Age mean±SD	65.9±3.5		64.6±14.4
Gender male/female	14/3		10/7
Histological types	Pooly differenciated SCC	8	Diffuse large B-cell lymphoma	14
	Moderate differenciated SCC	1	Adult T-cell lymphoma	2
	Nonkeratinizing differenciated SCC	1	angioimmunoblastic T-cell lymphoma	1
	Sarcomatoid SCC	1
	SCC;Undefined	6
Clinical stage			I	0
			II	3
			III	1
			IV	13

SD: standard deviation

CUP: cancer of unknown primary

ML: malignant lymphoma

SCC: squamous cell carcinoma

Comparisons of SUVmax and SUVmean between cervical LN metastasis from CUP and cervical LN involvement of ML

Table 2 shows the SUVmax and SUVmean measurements. The differences in SUVmax and SUVmean between cervical LN metastasis from CUP and cervical LN involvement of ML were not statistically significant.

Table 2

Comparisons of SUVmax and SUVmean between cervical LN metastasis from CUP and cervical LN involvement of ML

	Cervical LN metastasis from CUP		Cervical LN involvement from ML
	Mean	SD	Mean	SD	p
SUVmax	16.01	4.42	17.06	9.43	0.718
SUVmean	9.72	3.19	10.43	5.84	0.877

LN: lymph node

SUV: standardized uptake value

CUP: cancer of unknown primary

ML: malignant lymphoma

SD: standard deviation

Comparisons of texture features between cervical LN metastasis from CUP and cervical LN involvement of ML

Measurements of 34 texture features and correlations between tumor voxel and texture features are shown in Appendix 2 and 3, respectively. Texture features that had a strong correlation coefficient of >0.7 between tumor voxel were excluded to avoid influence of the confounding factor.[32]

Table 3 summarize the p-value, sensitivity, specificity, accuracy, and AUC of the selected texture features. Significant differences in 9 texture features in the total CSA that discriminated cervical LN metastases from CUP and cervical LN involvement from ML were observed. The highest AUC in the total CSA, which were obtained from the correlation in GLCM, were 0.851, respectively.

Table 3

AUC for selected texture features in the maximum cross-sectional area to discriminate between cervical LN metastasis from CUP and cervical LN involvement of ML

	Cervical LN metastasis from CUP		Cervical LN involvement of ML
	Mean	SD	Mean	SD	p	cut- off	SEN (%)	SPE (%)	ACC (%)	AUC
Histogram
Kurtosis	3.634	0.588	3.108	0.251	0.006*	3.24	82.4	76.5	79.4	0.844
Entropy	0.851	0.069	0.782	0.092	0.012*	0.82	70.6	70.6	70.6	0.753
Energy	0.169	0.029	0.195	0.041	0.020*	0.18	64.7	70.6	67.6	0.732
GLCM
Energy	0.031	0.012	0.042	0.019	0.025*	0.043	52.9	94.1	73.5	0.725
Correlation	0.193	0.074	0.092	0.058	0.001*	0.105	76.5	94.1	85.3	0.851
Entropy	1.673	0.142	1.538	0.185	0.019*	1.500	47.1	94.1	70.6	0.734
NGLDM
Coarseness	0.002	0.005	0.009	0.012	0.001*	0.004	62.7	88.2	79.4	0.830
GLZLM
SZE	0.686	0.019	0.635	0.083	0.048*	0.665	58.8	82.3	73.5	0.689
HGZE	10874	119	10966	117	0.037*	11030	35.3	100	67.6	0.709

CUP: cancer of unknown primary

ML: malignant lymphoma

SD: standard deviation

SEN: sensitivity

SPE: specificity

ACC: accuracy

AUC: the area under the curve

GLCM: gray-level co-occurrence matrix

NGLDM: neighborhood gray-level different matrix

GLZLM: gray-level zone length matrix

SZE: short-zone emphasis

HGRE: high gray-level run emphasis

* indicates significant differences.

The highest AUC and accuracy by SVM were 0.930 and 84.8%, respectively, with a combination of the kurtosis in Histogram, the correlation in GLCM, and the coarseness in the NGLDM as shown in Table 4 and Fig 2. Fig 3 show representative cases in which differences between cervical LN metastasis from CUP and cervical LN involvement of ML were identified using the combination of texture features.

Table 4

The best AUC for the combination of texture features using SVM

	SEN (%)	SPE (%)	ACC (%)	AUC
Kurtosis in Histogram, Correlation in GLCM, and Coarseness in NGLDM	82.6	98.9	84.8	0.930

SEN: sensitivity

SPE: specificity

ACC: accuracy

AUC: the area under the curve

SVM: support vector machine

GLCM: gray-level co-occurrence matrix features

NGLDM: neighborhood gray-level different matrix

Fig. 2

The ROC curves of the best combination of selected texture features

Fig. 3

A 70-year-old female with cervical LN involvement from ML

The correlation in the GLCM (0.018; cut-off values <0.110), the coarseness in the NGLDM (0.036; cut-off values >0.002), the kurtosis in the histogram (2.490; cut-off values <3.290), all derived from the total cross-sectional area, revealed true positives, while the SUVmax and SUVmean values were 9.88 and 6.02, respectively (A and B).

GLCM: gray-level co-occurrence matrix features

NGLDM: neighborhood gray-level different matrix

SZE: short-zone emphasis

GLZLM: gray-level zone length matrix

DISCUSSION

This study demonstrates that 9 texture features of the total CSA of the enlarged cervical LNs on unenhanced CT, differentiated cervical LN metastases with CUP versus cervical LN involvement from ML; however, no significant differences in the SUVmax or SUVmean were observed. Thus, it can be considered that radiomics analysis might provide useful information for the differentiation of cervical lymphadenopathy between CUP and ML.

Head-and-neck CT for cervical lymphadenopathy has been performed to evaluate the location, abnormal internal architecture, extra-nodal metastasis, and primary site of malignant tumors. Notably, the necrotic components that have a key role in the evaluation of malignancy are frequently estimated by the enhancement pattern on a head-and-neck CT with contrast media.[33,34] On unenhanced CT, it is difficult to detect small changes in CT attenuation objectively; such changes are likely to be diagnosed based on radiologists’ impression. Several previous studies have demonstrated that texture analysis can be used to evaluate correlations between imaging findings on unenhanced CT and pathological findings and to differentiate between malignant and benign tumors.[35-37] Cell proliferation, myxoid changes, abnormal angiogenesis, and necrotic changes within malignant tumors result in heterogeneity.[37] Central necrosis in cervical LNs that reflects reduced CT attenuation is recognized as indicating metastasis.[36,39] Furthermore, ML appears as isoattenuation in a homogenous mass.[39] Nodal lymphoma is characterized by high cellularity, large nuclei, and less extracellular space than well or moderate differentiated carcinomas.[34,41] In the present study, unenhanced CT-based quantitative analysis of texture features could be used to differentiate between cervical LN metastases from CUP and cervical LN involvement from ML. Texture features related to randomness, such as entropy in the GLCM, in cases with cervical LN metastases from CUP were higher than those with cervical LN involvement from ML and the coarseness in the NGLDM related to homogeneity was lower. Some texture features could depict the gray-level differences between cervical LN metastasis from CUP and cervical LN involvement of ML on unenhanced CT.

Unenhanced CT was performed with a low radiation dose in the current study. Reduced radiation doses lead to decreased image quality due to a decrease in photons. However, unenhanced low-dose CT might be useful for quantifying enlarged cervical LN. A previous study reported that CT-based texture analysis is not affected by changes in tube current on CT in a phantom study; thus, findings support the present results.[42]

This study has some limitations. First, we had a small number of patients in this retrospective and single institutional study. Second, occult EBV-related nasopharyngeal cancers and occult HPV-related oropharyngeal cancers were not excluded due to the inclusion of patients who were treated before the recommendation to evaluate EBV and HPV infection status was implemented. Third, the ROIs for the cervical LNs were manually delineated. Fourth, the degree of aggressiveness of each ML was not evaluated. Ganshan et al suggested the possibility that CT-based texture analysis for ML may be correlated with fibrosis, which appears in a variety of patterns in ML subtypes.[42] Necrosis within adult T-cell lymphoma lesions has previously been described to be correlated with poor prognosis.[43] The aggressiveness of ML can affect texture analysis. Therefore, further large-scale studies that evaluate more parameters and classify ML based on the degree of aggressiveness are required.

In conclusion, CT-based texture analysis can distinguish cervical LN metastasis from CUP and enlarged cervical LNs in ML, while the SUVmax and SUVmean cannot differentiate them. Quantitative analysis of texture features on unenhanced CT has the potential to provide additional information about patients with malignant cervical lymphadenopathy.

CONFLICT OF INTEREST

The authors have declared no conflicts of interest.

Click here for additional data file.

Comparisons of texture features in the total cross-sectional area between cervical LN metastasis from CUP and cervical LN involvement of ML

Click here for additional data file.

Correlations between texture features and tumor voxels

Click here for additional data file.