CT texture analysis can help differentiate between malignant and benign lymph nodes in the mediastinum in patients suspected for lung cancer

Abstract

Background

In patients with non-small-cell lung carcinoma NSCLC the lymph node staging in the mediastinum is important due to impact on management and prognosis. Computed tomography texture analysis (CTTA) is a postprocessing technique that can evaluate the heterogeneity of marked regions in images.

Purpose

To evaluate if CTTA can differentiate between malignant and benign lymph nodes in a cohort of patients with suspected lung cancer.

Material and Methods

With tissue sampling as reference standard, 46 lymph nodes from 29 patients were analyzed using CTTA. For each lymph node, CTTA was performed using a research software “TexRAD” by drawing a region of interest (ROI) on all available axial contrast-enhanced computed tomography (CT) slices covering the entire volume of the lymph node. Lymph node CTTA comprised image filtration-histogram analysis undertakes two stages: the first step comprised an application of a Laplacian of Gaussian filter to highlight fine to coarse textures within the ROI, followed by a quantification of textures via histogram analysis using mean gray-level intensity from the entire volume of the lymph nodes.

Results

CTTA demonstrated a statistically significant difference between the malignant and the benign lymph nodes (P = 0.001), and by binary logistic regression we obtained a sensitivity of 53% and specificity of 97% in the test population. The area under the receiver operating curve was 83.4% and reproducibility was excellent.

Conclusion

CTTA may be helpful in differentiating between malignant and benign lymph nodes in the mediastinum in patients suspected for lung cancer, with a low intra-observer variance.

Keywords

Computed tomography (CT)computer applications – detection/diagnosis thorax lymphatic mediastinum

Introduction

In patients with suspected lung cancer, contrast enhanced computed tomography (CT) and/or fluorine-18 fluoro-D-glucose positron emission tomography/computed tomography (F18-FDG PET/CT) are standard procedures in the identification, characterization, and staging of the disease. CT relies on size and contrast enhancement, and PET/CT on the metabolism of F18-FDG. Studies have shown, that although CT and F18-FDG PET/CT both improve the diagnostic accuracy, invasive procedures are still mandatory to confirm/exclude malignancy (1,2). Improving the diagnostic accuracy of imaging modalities to identify malignant involvement of the mediastinal lymph nodes in a patient with a suspected lung nodule/mass is important in the diagnostic work-up when therapy and prognosis is considered (3 –5). This will further assist in minimizing the number of invasive procedures required to confirm the presence of malignancy within lymph nodes and therefore the complications associated with the invasive procedures.

Contrast-enhanced CT texture analysis (CTTA), a new postprocessing technique, has been successfully applied in different oncological applications, to non-invasively quantify heterogeneity within different tumors and potentially leading to improved prognosis, disease risk-stratification/characterization, and treatment-response/prediction in a number of cancer sites, e.g. non-small-cell lung cancer (NSCLC) (6,7), esophageal (8), colorectal (primary and metastatic liver) (9), head and neck (10), and metastatic renal cell cancer (11). Additional texture features on CT seem to be associated with tumor biology (providing a biological-correlate) such as grade, physiology (perfusion and glucose uptake), hypoxia, angiogenesis, and genetic mutation in different tumors such as gliomas, colorectal, NSCLC, and esophageal cancer (8,9,12).

The purpose of this study was to evaluate if CTTA can differentiate between malignant and benign lymph nodes in the mediastinum in patients suspected to have lung cancer with tissue sampling as gold standard.

Material and Methods

The study conformed to Danish legal requirements. As all subjects received best patient care, Institutional Review Board approval was waived (Journal number: 1-15-0-72-2-09).

Patients

During a 15-month study period, 840 patients with suspected lung cancer were referred to a tertiary hospital. All patients had a CT performed and subjected to a multidisciplinary team (MDT) decision. When indeterminate lesions or malignancy suspect lesions were present on CT, the patients received an additional F18-FDG PET/CT. In total, 168 patients received both CT and an F18-FDG PET/CT during the study period. Among the patients 32 had NSCLC and 29 patients were excluded from analysis because the CT scan was performed at other institutions using other CT scanners and contrast media policy. Thus, the study population consisted of 107 patients. A potential number of 1926 lymph node stations were available for analysis.

The inclusion criteria for a lymph node was: (i) a long axis diameter of >1 cm as to reliably get some better statistics in terms of more number of voxels available within the lymph node for CTTA calculations; (ii) the largest lymph node within each lymph node station was selected; (iii) lymph nodes with large vessels going through was excluded as such vessels may disturb the texture calculations; (iv) lymph nodes distorted by artifacts were excluded as such artifacts may also disturb texture calculations; and (v) tissue sampling was performed either by surgery or mediastinoscopy.

A total of 1788 lymph node stations did not contain a lymph node with a long axis diameter of >1 cm; 38 lymph node stations were hampered by CT artifacts; 30 lymph node stations were in patients with distant metastases so no histopathology was obtained, 17 lymph node stations were out of the range of mediastinoscopy or surgery, and in seven lymph node stations tissue sampling was not performed for unknown reasons. So 46 lymph nodes in 29 patients were appropriate for analysis (Fig. 1).

Fig. 1.

Flow chart for obtaining final study group.

Contrast-enhanced CT

CT examinations included the chest and the upper abdomen. CT was performed with a multiple-row detector CT scanner (Philips Brilliance CT 64-channel scanner; Philips Healthcare, Best, The Netherlands). CT acquisition parameters were: collimation, 64 × 0.625 mm; kV, 120–140; mAs/slice, 150–250; rotation time, 0.75; reconstruction thickness, 2 mm; increment, 1 mm; pitch, 1.078; field of view, 35 cm; and matrix, 512 × 512. CT examinations included the chest and the upper abdomen. Iodixanol 270 mg/mL (Visipaque® 270; GE Healthcare, Oslo, Norway), or iohexol 300 mg/mL (Omnipaque® 300; GE Healthcare), was injected intravenously in weight-adjusted doses of 2 mL/kg body weight to compensate for differences in distribution volume. A bolus tracking technique with a region of interest (ROI) in the descending aorta on the level of carina was used to compensate for differences in cardiac output. CT was performed after a delay of 15 s for the chest and upper abdomen (late arterial phase), and 65 s for the upper abdomen (portovenous phase), and raw picture datasets were transferred to a Philips Extended Brilliance™ Workspace workstation v4.02 (Philips Healthcare) for further analysis (1,2).

Reference standard

Tissue sampling was obtained either by surgery with complete lymph node resection, by mediastinoscopy with sampling from nodal stations 2R/L, 4R/L, and 7 or by anterior mediastinotomy from stations 3A, 5, and 6. Among the 46 lymph nodes included in the study tissue sampling was obtained by surgery in 74% (34/46) and by mediastinoscopy/-tomy in 26% (12/46).

CTTA of lymph nodes

CTTA was performed using TexRAD, a research software (TexRAD Ltd, Cambridge, UK) The technique comprised an image filtration-histogram approach; in which an initial filtration step using a Laplacian of Gaussian band-pass filter (non-orthogonal Wavelet) was employed to extract and enhance features of different size based on the spatial scale filter (SSF) values varying from 1.4–4.1 mm in diameter (Fig. 2.) Following image filtration, each filtered image texture map was quantified using histogram parameters such as mean image intensity.

Fig. 2.

The CTTA window with a drawn ROI and highlighted features of fine (1.4 mm) –>coarse textures (4.1 mm). Fine textures in red, medium textures in green, and coarse textures in blue colors.

Two resident radiologists performed the analysis (over an interval of 3–6 months for the assessment of intra- and inter-observer variance). The radiologists were blinded to all patient identifiers and clinical data as well as results from CT, F18-FDG PET/CT, and tissue sampling. For each lymph node a ROI was drawn on all axial images. For conglomerates the entire structure was enclosed within a ROI. A semi-automated approach further refined the ROI, enclosing the lymph node to exclude air and fat, a thresholding procedure that removed pixels having value < –50 HU.

Statistical analysis

Mean gray-level intensity for all filter sizes was analyzed with the Shapiro-Wilk’s test as well as QQ-plots to determine if they were Gaussian distributed. All but the unfiltered mean image intensity showed a Gaussian distribution. Using the following transformation: $\sqrt[2]{(x_{max} + 1) - x}$ , where x = mean gray-level intensity and x_max = maximum mean gray-level intensity observed across all lesions. After transformation the unfiltered mean gray-level intensity approached Gaussian distribution for both the malignant and benign group.

The independent samples t-test was used to compare means between malignant and benign groups. In case of unequal variances determined by Levener’s test an unequal variance t-test would be performed instead (Welch’s t-test).

The variables that proved a statistically significantly difference between malignant and benign groups was used in a binary logistic regression model to obtain the 2 × 2 table used for calculating sensitivity, specificity, and accuracy. Furthermore, an overlay plot of sensitivity/specificity was performed to obtain the cutoff value giving the highest overall accuracy as well as to obtain the cutoff providing the highest specificity without losing sensitivity in order to avoid potential future patients to be deprived of possibly curative surgery. In case the two cutoff values were not identical the latter was used for analysis.

To assess the method a receiver operating curve (ROC) was created using the mentioned cutoff value.

Comparison of inter- and intra-observer variance was calculated using limits of agreement by a Bland-Altman plot (13).

Results

Lymph node morphology

Among the 46 lymph nodes available for analysis 15 was malignant and 31 benign evidenced by tissue sampling. The 15 malignant lymph nodes had a mean short axis diameter of 9.9 mm (range, 7.0–17.9 mm) and a mean long axis diameter of 16.3 mm (range, 10.1–21.6 mm) while the 31 benign lymph nodes had a mean short axis diameter of 9.8 mm (range, 6.5–14.4 mm) and a mean long axis diameter of 16.7 mm (range, 10.1–38.6 mm).

After transformation using the natural logarithm both the short axis diameter and long axis diameter were normally distributed for both malignant and benign lymph nodes as assessed by Shapiro-Wilk’s test (P > 0.5) as well as by QQ-plots.

There was no statistical significant difference for short axis (P = 0.882) and long axis diameter (P = 0.916) between the malignant and the benign groups as assessed by the independent t-test.

A short axis of 10 mm as a compromise is often considered as the threshold for pathologic lymph nodes in the mediastinum (14). In this study a correct identification of only seven of the 15 malignant lymph nodes and 19 of the 31 benign lymph nodes were obtained when a short axis larger than 10 mm was considered giving a sensitivity of 46.7% and specificity of 61.3%.

Malignant vs. benign lymph node textures

Mean gray-level intensity for all filter levels were normally distributed for both malignant and benign lymph nodes as assessed by Shapiro-Wilk’s test (P > 0.5) as well as visually by QQ-plots. This was obtained after transformation of the unfiltered mean image gray-level data as described in the statistical analysis section.

Only unfiltered mean image intensity (Mean HU) showed a statistically significant difference between the malignant and the benign group. The lymph nodes in the malignant group showed significantly higher mean image intensity than the benign group. Mean difference was 1.88 (95% CI, 0.88–2.88), as Levener’s test showed unequal variances (P = 0.029) the results presented are from the unequal variance t-test (Welch’s t-test) performed on the transformed data: t(19.556) = 3.939, P = 0.001. Mean image intensity at the coarse filtered texture scale (SSF = 3.4 mm) showed a trend to differentiate between the malignant and benign group but did not reach statistical significance. Mean difference was −6.59 (95% CI, −13.62–0.45) and the result of the independent t-test was t(44) = −1.888, P = 0.066 for filter size 3.4 mm. No other texture parameters showed significant differences between the two groups.

In Table 1 the distribution characteristics and P values for the texture mean image intensity is shown and Fig. 3 shows the box and whisker plot for the unfiltered mean image intensity. Three outliers were found, however, when performing both a Mann-Whitney U test and the unequal variance t-test they both gave roughly the same result, hence the outliers were included in the analysis.

Table 1.

Distribution characteristics for CTTA (mean image intensity) stratified for benign and malignant lymph nodes.

Parameter	Mean	95% CI	P values
	CTTA benign (mean image intensity)
Without filtration	55.66	50.36–60.98	0.001
Filter = 1.4 mm	3.22	0.63–5.82	0.566
Filter = 2.0 mm	−8.63	−13.12– −4.14	0.479
Filter = 2.7 mm	−14.03	−18.15– −9.90	0.152
Filter = 3.4 mm	−15.06	−18.77– −11.35	0.066
Filter = 4.1 mm	−11.73	−15.71– −7.75	0.366
	CTTA malignant (Mean image intensity
Without filtration	75.32	70.23–80.42
Filter = 1.4 mm	2.4	−0.09–4.90
Filter = 2.0 mm	−5.17	−8.90– −1.44
Filter = 2.7 mm	−7.97	−11.67– −4.27
Filter = 3.4 mm	−8.26	−12.60– −3.92
Filter = 4.1 mm	−8.24	−12.94– −3.54

P values for the independent t-test with a null hypothesis that there is no difference between the two groups are listed as well.

Fig. 3.

Box-and-whisker plot for unfiltered mean divided between benign (1) and malignant (2) group. Notice the three outliers in the benign group.

Binary logistic regression analysis and ROC curve

The binary logistic regression analysis on the unfiltered mean image intensity provided a classification table for the CTTA as well as a way to obtain cutoff values for future studies. In order to determine the optimal cutoff value for both sensitivity and specificity we performed an overlay plot of sensitivity and specificity at various probability cutoffs (Fig. 4).

Fig. 4.

Sensitivity/specificity plotted against cutoff values.

The logistic regression was performed to assess the effects of unfiltered mean values on the likelihood of a lymph node to be malignant. We used a cutoff of 0.6 instead of the intersection value to obtain the highest value of specificity without losing too much sensitivity to avoid patients to be deprived of possible curative surgery (Table 2). The logistic regression model was statistically significant, χ²(1) = 18.014, P < 0.0005. The model demonstrated 45.2% (Nagelkerke R²) of the variance in the malignant lymph nodes and classified 82.6% of the cases correctly (sensitivity = 53.3%, specificity = 96.8%, positive predictive value = 88.9%, and negative predictive value = 81.1%).

Table 2.

2 × 2 classification table derived from the binary logistic regression analysis.

Cutoff = 0.6	Benign	Malign
Malignant	1	8
Benign	30	7

Furthermore, a ROC curve was derived from the data provided by the logistic regression analysis which gave us an area under the curve of 83.4% with a 95% confidence interval of 70.9–96.0% (Fig. 5).

Fig. 5.

Receiver operating curve for the CTTA technique. Area under the curve = 83.4% and P < 0.0001.

Intra- and inter-observer variance

The median ratio of reproducibility for intra-observer variance was essentially zero. Thus, no systematic bias was observed between the two readings. Furthermore, the limits of agreement were narrow. The mean difference between malignant and benign lymph nodes was 19.97 HU units and the limits of agreement were in the range of −3.94 to 3.72 HU units. The variability was roughly consistent across the graph (Fig. 6.).

Fig. 6.

Agreement plots for unfiltered mean intra-observer variance. Punctuated line show the 95% limits of agreement while the solid line show the mean of differences.

However, with regard to inter-observer variance a systematic bias was observed, as the mean of the difference was 7.19 HU units. Furthermore, the limits of agreement were wide (Fig. 7.).

Fig. 7.

Agreement plots for unfiltered mean inter-observer variance. Punctuated line show the 95% limits of agreement while the solid line show the mean of differences. Notice the mean of differences suggest a systematic bias.

Discussion

In a previous study (2,15), both CT as well as F18-FDG PET/CT performed equally poor when mediastinal disease in lung cancer patients were considered Furthermore, in the presented study, the mean size of the malignant lymph nodes were not significantly larger than the benign. Of 46 lymph nodes, 27 were not pathologically enlarged on CT (larger than 10 mm in the short axis) (14). In RECIST 1.1 (16) the short axis needs to be larger than 15 mm before a lymph node is considered pathologic. F18-FDG PET/CT has in several studies demonstrated difficulties when assessing small lesions (1,2,17), so most likely many of the lymph nodes in our study would be below the detection limit for F18-FDG PET/CT.

Consequently, at most institutions even non-enlarged lymph nodes and nodes without significant uptake of FDG go for endobronchial ultrasound/endoscopic ultrasound (EBUS/EUS). Unfortunately, this study was performed at a time where mediastinoscopy was standard procedure at our institution. Although EBUS/EUS can add to the diagnostic accuracy in these small lesions the accuracy is not 100% (18,19). In a recent report (20) the authors also found that texture analysis may be complementary in the CT evaluation of mediastinal lymph nodes in lung cancer patients. In that study, non-enhanced CT scans were considered. So although from a theoretical point of view differences in heterogeneity is expected to be augmented in a contrast enhanced phase it also seems that a non-enhanced CT can display important heterogeneity features. Further studies comparing the impact of texture features in different contrast media phases is warranted.

In the present study, we found eight malignant non-enlarged lymph nodes and 12 enlarged benign lymph nodes in accordance to the gold standard. Additional texture analysis found five malignant non-enlarged lymph nodes and benignancy in 11 enlarged lymph nodes. Thus in this study with 33% prevalence of malignancy we had a sensitivity of 53.3% and a specificity of 96.8%. Thus, on a per lesion basis CTTA and CT is supplementary when considering both malignancy and benignancy.

The unfiltered mean image gray level showed a significant difference between malignant and benign lesions. However, when applying different filter sizes it showed a strong tendency for a filter size of 3.4 mm (P = 0.066) although non-significant to distinguish between malignant and benign lymph nodes. Probably the low overall number of lymph nodes analyzed as well as the small size of lymph nodes is the reason for the non-significant results. CTTA has turned out to be a useful postprocessing modality and have been used for prognosis, disease risk-stratification/characterization and treatment-response/prediction, as well as been correlated with physiology (perfusion and glucose uptake), hypoxia, angiogenesis, and genetic mutation in different tumors (6 –9,11,12).

CTTA is highly reproducible when intra-observer variance is considered. However, a systematic bias was introduced when looking at inter-observer variance. The second reader consistently had higher values than the first reader. This may reflect inclusion of nearby calcifications or vessel walls. Despite this bias, it had no impact on the differentiation between malignant and benign lymph nodes.

Forty-six lymph nodes were biopsied and found suitable for analysis, of these 67% were benign. As the prevalence of lung cancer was 80% in patients that had both FDG-PET/CT and CT, and as metastatic disease in patients with malignant lymph nodes often was present in this cohort, tissue sampling was only obtained in 46 lymph nodes that also met the other inclusion criteria. Thus, a selection bias was introduced, as mediastinal staging was not necessary in many of the patients. Even though the lymph nodes examined by CTTA are subjected to a selection bias, the study population reflects the challenges experienced in an MDT setting of a tertiary lung cancer center. Despite the selection bias, CTTA seems to add complementary diagnostic information as 16 of the 46 lymph nodes changed status from malignant to benign or vice versa in accordance with the gold standard, when compared to normal CT.

The results presented in this paper could potentially open up for characterizing lymph nodes in other cancer forms as well as it would be plausible to believe similar changes would be seen in those settings as well.

The study has limitations. First, a relative small number of patients were included. Second, our selection bias also limited our study. The reason for this bias was to ensure that different CT techniques with different acquisition parameters and different contrast media policies was not reflected in the heterogeneity measures of texture analysis. Certainly if the study was performed today, more lymph nodes could have been included due to the use of EBUS/EUS guided biopsies in a routine setting. However, the strict selection of lymph nodes evaluated is also a strength when the impact of texture analysis of lymph nodes is considered. In future studies the introduction of EBUS/EUS and CT guided biopsies from lymph nodes in all lymph node stations may ensure a “full” N-staging so the N-staging by texture can be evaluated on a per patient basis instead of a per lesion basis. Whenever an imaging technique is compared with histopathology of a surgical specimen there will always be the risk that it is not the same lesion that is compared. The risk for such mistakes in the present study is reduced as only the largest lymph node in a given lymph node station on CT was compared with the largest lymph node in the corresponding surgical specimen.

In conclusion, CTTA in mediastinal lymph nodes may be helpful and add information to CT in the differentiations between metastatic and benign lymph nodes in patients suspected for lung cancer. However, the technique does not yet have the diagnostic impact necessary to replace current diagnostic measures. Further studies comparing CT + CTTA to EBUS/EUS and F18-FDG PET/CT on a patient-by-patient basis is warranted.

Footnotes

Conflict of interest

BG is a part-time employee and board member of TexRAD Ltd.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

References

Harders

Madsen

Hjorthaug

. Characterization of pulmonary lesions in patients with suspected lung cancer: computed tomography versus [¹⁸F] fluorodeoxyglucose-positron emission tomography/computed tomography. Cancer Imaging 2012; 12: 437–446.

Harders

Madsen

Hjorthaug

. Mediastinal staging in non-small-cell lung carcinoma: computed tomography versus F-18-fluorodeoxyglucose positron-emission tomography and computed tomography. Cancer Imaging 2014; 14: 23–23.

Silvestri

Gonzalez

Jantz

. Methods for staging non-small cell lung cancer: Diagnosis and management of lung cancer, 3rd ed: American College of Chest Physicians evidence-based clinical practice guidelines. Chest 2013; 143: e211S–250S.

Darling

Maziak

Inculet

. Positron emission tomography-computed tomography compared with invasive mediastinal staging in non-small cell lung cancer: results of mediastinal staging in the early lung positron emission tomography trial. J Thorac Oncol 2011; 6: 1367–1372.

Y-L

Yuan

D-M

Wang

. Diagnostic performance of integrated positron emission tomography/computed tomography for mediastinal lymph node staging in non-small cell lung cancer: a bivariate systematic review and meta-analysis. J Thorac Oncol 2011; 6: 1350–1358.

Ganeshan

Abaleke

Young

RCD

. Texture analysis of non-small cell lung cancer on unenhanced computed tomography: initial evidence for a relationship with tumour glucose metabolism and stage. Cancer Imaging 2010; 10: 137–143.

Ganeshan

Panayiotou

Burnand

. Tumour heterogeneity in non-small cell lung carcinoma assessed by CT texture analysis: a potential marker of survival. Eur Radiol 2012; 22: 796–802.

Ganeshan

Skogen

Pressney

. Tumour heterogeneity in oesophageal cancer assessed by CT texture analysis: preliminary evidence of an association with tumour metabolism, stage, and survival. Clin Radiol 2012; 67: 157–164.

Ganeshan

Kozarski

. Assessment of primary colorectal cancer heterogeneity by using whole-tumor texture analysis: contrast-enhanced CT texture as a biomarker of 5-year survival. Radiology 2013; 266: 177–184.

10.

Zhang

Graham

Elci

. Locally advanced squamous cell carcinoma of the head and neck: CT texture and histogram analysis allow independent prediction of overall survival in patients treated with induction chemotherapy. Radiology 2013; 269: 801–809.

11.

Goh

Ganeshan

Nathan

. Assessment of response to tyrosine kinase inhibitors in metastatic renal cell cancer: CT texture as a predictive biomarker. Radiology 2011; 261: 165–171.

12.

Skogen

Ganeshan

Good

. Measurements of heterogeneity in gliomas on computed tomography relationship to tumour grade. J Neurooncol 2013; 111: 213–219.

13.

Bland

Altman

. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986 8; 1: 307–310.

14.

Prenzel

Mönig

Sinning

. Lymph node size and metastatic infiltration in non-small cell lung cancer. Chest 2003; 123: 463–467.

15.

Gómez-Caro

Boada

Cabañas

. False-negative rate after positron emission tomography/computer tomography scan for mediastinal staging in cI stage non-small-cell lung cancer. Eur J Cardiothorac Surg 2012; 42: 93–100discussion 100.

16.

Eisenhauer

Therasse

Bogaerts

. New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1). Eur J Cancer 2009; 45: 228–247.

17.

Carrillo

Daniel

Hall

. Fusion positron emission/computed tomography underestimates the presence of hilar nodal metastases in patients with resected non-small cell lung cancer. Ann Thorac Surg 2012; 93: 1621–1624.

18.

Fernández-Esparrach

Sendino

. [New endoscopic ultrasound (EUS) applications in lung cancer: evaluation of patients with negative mediastinal CT and re-staging after neoadjuvant treatment]. Arch Bronconeumol 2011; 47: 410–414.

19.

Herth

FJF

. Nonsurgical staging of the mediastinum: EBUS and EUS. Semin Respir Crit Care Med 2011; 32: 62–68.

20.

Bayanati

E Thornhill

Souza

. Quantitative CT texture and shape analysis: can it differentiate benign and malignant mediastinal lymph nodes in patients with primary lung cancer? Eur Radiol 2015; 25: 480–487.