Abstract
Background/Aim: Texture analysis can provide quantitative imaging markers from computed tomography (CT) images. The Node-RADS classification was recently published as a classification system to better characterize lymph nodes in oncological imaging. The present analysis investigated the diagnostic benefit of CT texture analysis and the Node-RADS classification to categorize and stage lymph nodes in patients with perihilar cholangiocarcinoma. Patients and Methods: Overall, 25 patients (n=9 females, 36%) with a mean age of 72.4±8.1 years were included. All patients were surgically resected and the lymph nodes were histopathologically analyzed. CT-texture analysis was performed with the Mazda package. All investigated lymph nodes were scored in accordance with the Node-RADS classification. Results: Regarding lymph node discrimination (N− versus N+), Node-RADS classification achieved an area under the curve (AUC) of 0.86 resulting in a sensitivity of 78% and a specificity of 86%. Multiple investigated texture features were different between negative and positive lymph nodes. The “S(0,1)SumVarnc” achieved the best AUC of 0.75 resulting in a sensitivity of 0.91 and a specificity of 0.67. Correlation analysis showed various statistically significant associations between CT texture features and Node-RADS score. Conclusion: Several CT texture features and the Node-RADS score derived from preoperative staging CT were associated with the malignancy of the hilar lymph nodes and might aid for preoperative staging. This could change surgical treatment planning in hilar cholangiocarcinoma.
Perihilar cholangiocarcinoma (pCCA) is a malignant tumor of the bifurcation of the hepatic duct, which is frequently diagnosed at an advanced stage with consequently poor prognosis (1-4). Surgical treatment is still considered as the only treatment with curative intention (1-3). Thus, preoperative correct staging can have a substantial impact on potential treatment decisions.
In clinical routine, ultrasonography, computed tomography (CT) and magnetic resonance imaging (MRI) have all been used and investigated for pCCA diagnosis and staging purposes (1, 2, 5, 6).
However, imaging is challenging for this tumor entity due to small tumor size and complex surrounding anatomic structures. That is why contrast enhanced CT imaging has a tumor detection rate of only 60-69% and accuracy in determining resectability of 44-80% (7). Regarding correct nodal staging, CT imaging was reported to have low accuracy. A meta-analysis by Ruys et al. provided an overall pooled sensitivity of 61% and specificity of 88% to diagnose nodal metastases (8). Moreover, 13% of lymph nodes smaller than 10 mm in short axis diameter were staged as a metastatic lymph node, and only 25% of nodes larger than 10 mm actually harboured malignant cells (7, 9).
Texture analysis on CT images can be used to provide a quantitative imaging marker (10-12). Various emergent benefits of CT texture analysis have been investigated in recent years, mainly in the diagnostic field of oncology (10-12). Employing this analysis, different spatial characteristics of CT images of the analyzed tumors can be utilized for diagnostic and discrimination purposes (10-12). This technique was also previously investigated to further characterize benign and malignant lymph nodes in different localizations of the body (13-15).
Node-Reporting and Data System (Node-RADS) was recently published as a promising classification system to standardize the categorization for lymph nodes in clinical staging (16). It consists of two main features, namely “size” and “configuration”. It uses a 5-point probability scale ranging from 1 with very low likelihood for malignancy up to 5 with very high likelihood (16). Yet, this classification system was not systematically investigated in clinical cohorts. For other standardized radiological reporting systems, comprising breast imaging RADS (BI-RADS), prostate imaging RADS (PI-RADS), liver imaging RADS (LI-RADs), thyroid imaging RADS (TI-RADS), various studies have analyzed the potential diagnostic benefits and potential shortcomings as well as an evaluation of malignancy probability in the different categories (17-20).
However, there is little data regarding the diagnostic benefit of Node-RADS beyond the first description (16, 21). Moreover, despite promising published results for texture analysis to better characterize lymph nodes, no study investigated this imaging analysis for lymph nodes in patients with pCCA.
Therefore, the purpose of the present study was to elucidate whether CT-derived texture analysis parameters and Node-RADS can stratify hilar liver lymph nodes and to improve the diagnostic performance for discrimination purposes in patients with pCCA.
Patients and Methods
Study design. This retrospective, observational study involving human participants was performed in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. It received ethical approval from the local ethics committee at the Medical Faculty (EK: 243-14-1407-2014).
The radiological database of one university hospital was retrospectively screened for patients with pCCA between January 2016 and December 2021.
Inclusion criteria consisted of available presurgical CT images of histopathologically confirmed pCCA. All analyzed CT scans were acquired within one month before the surgical procedure. Overall, 25 patients (n=9 females, 36%) with a mean age of 72.4±8.1 years were included. An overview of the descriptive statistics of the included patients is given in Table I.
Demographic characteristics of the investigated lymph nodes.
Surgery and pathology. The Brisbane 2000 classification system was used to describe hepatic anatomy and resection (22). The patients underwent anatomic left or right hemihepatectomies with or without wedge resections of adjacent liver segments or extended left and right hemihepatectomies commonly entitled as trisectionectomy. The exact method of resection depended on the localization of the tumor, its extension and the depth of the infiltration. All procedures were classified in accordance to the Bismuth–Corlette classification (23, 24). In all cases, the tumor resection was combined with a systematic lymph node dissection of the liver hilum. According to clinical routine all suspicious lymph nodes were removed and analyzed by histology.
Imaging technique. In all cases, CT was performed with a 128-slice CT scanner (Ingenuity 128, Philips, Hamburg, Germany) during clinical work up. For every patient, intravenous administration of an iodine-based contrast medium (90 ml Imeron 400 MCT, Bracco Imaging Germany GmbH, Konstanz, Germany) was given at a rate of 2-4.0 ml/s via a peripheral venous line. All investigated CT images were obtained in portalvenous phase after 70 s. Automatic bolus tracking was performed in the aorta descenders with a trigger of 100 Hounsfield units (HU). Typical imaging parameters were: 100 kVp; 125 mAs; slice thickness,1 mm; pitch, 0.9.
Texture analysis. CT images were further processed with the dedicated software MaZda (version 4.7) (25, 26). A polygonal region of interest (ROI) was placed on the largest, representative CT slice of the hilar lymph node. The ROI placement was performed in a blinded manner to the clinical and histopathology results. Gray-level (μ) normalization was performed to reduce contrast and brightness variation employing the limitation of dynamics to μ±3 standard deviations as performed in previous analyses (21, 27). The extracted imaging parameters were in the groups of first order histogram features, second order texture features comprising gray-level histogram co-occurrence matrix run-length matrix, autoregressive model (theta 1 to 4, sigma), and wavelet transform. In total, 279 texture features were retrieved for every patient and analyzed lymph node. Figure 1 displays two representative cases of the patient sample for illustration purposes.
Representative cases of patient samples. The region of interest for the texture analysis is shown in red. A) pT2a, N0, M0, Bismuth 4, Node-RADS 1, short-axis-diameter 7 mm. B) pT2b, N1, M0, Bismuth 4, Node-RADS 3, short-axis-diameter 9 mm. C) pT2b, N1, M0, Bismuth 4, G2, Node-RADS 4, short-axis-diameter 12 mm.
Lymph node score. The lymph nodes located in the liver hilum were scored according to the previously described Node-RADS classification (16). Scoring was performed by two radiologists independently with 4 years and 2 years of experience in oncological CT imaging analysis, respectively, both blinded to the pathological results. The scoring system ranges from 1 to 5 reflecting the level of probability of malignancy: “1–very low”; “2–low”; “3–equivocal”; “4–high”; “5–very high, as in the other RADS-classifications (17-20). The size and configuration are considered as the two main imaging findings. The size was classified as enlarged, when short axis diameter was above 10 mm. Regarding configuration, the CT texture was assessed in a qualitative manner and categorized as homogenous, heterogeneous, focal or gross necrosis. The lymph node border can be described as smooth or irregular. Shape was defined as kidney bean with fat hilus or spherical without fat hilus. All features together resulted in the final the lymph node category.
Statistical analysis. The statistical analysis and graphics creation were performed with SPSS (IBM, Version 25.0; Armonk, NY, USA). The collected data was first analyzed with descriptive statistics. The associations between pathological outcome and the imaging findings were analyzed with Spearman’s correlation coefficient. Discrimination analysis was performed with Mann–Whitney test for continuous data and Fisher’s exact test for categorical data. Receiver-operating characteristics (ROC) curve analysis was used to test for diagnostic accuracy with area under the curve (AUC)-analysis. Interreader agreement was assessed with Cohen’s kappa. In all instances, two-sided p-values <0.05 indicated statistical significance.
Results
Table I provides the demographics of the included patient sample.
Discrimination analysis for node-RADS score according to nodal status. Node-RADS-scoring resulted for reader 1 in a total of n=15 for Node-RADS 1 (30%), n=14 for Node-RADS 2 (28%), n=15 for Node-RADS 3 (30%), n=5 for Node-RADS 4 (10%) and n=1 (2%) for Node-RADS 5. For reader 2, the results were n=18 for Node-RADS 1 (36%), n=12 for Node-RADS 2 (24%), n=6 for Node-RADS 3 (12%), n=10 for Node-RADS 4 (20%) and n=4 (8%) for Node-RADS 5. Inter-reader agreement was only fair for both readers (k=0.35, p<0.001). The distribution of malignant lymph nodes is shown Table II in accordance to the Node-RADS score.
Distribution of malignancy according to each Node-RADS score for both readers.
For reader 1, Node-RADS 1 had a malignancy rate of 13.3% and for reader 2 0%, Node-RADS 2 had a malignancy rate of 15.4% for reader 1 and of 25 % for reader 2, Node-RADS 3 had a malignancy rate of 80% for reader 1 and 75% for reader 2, Node-RADS 4 and 5 both had 100% for both readers. Node-RADS score showed statistically significant differences between N0 and N1-3 stage (for reader 1: mean 1.63±0.68 score for N0 versus 3.00±0.95 score for N1-3, for reader 2: 1.33±0.48 score for N0 versus 3.65±0.94 score for N1-3, p<0.001, respectively).
The corresponding ROC curve analysis achieved an AUC of 0.86. Employing the threshold value of Node-RADS 2, resulted in a sensitivity of 0.78 and a specificity of 0.86 (Figure 2). Short axis diameter reached statistically significant difference between N0 and N1-3 stage (mean 6.2±1.3 mm for N0 versus 8.6±1.7 mm for N1-3, p<0.001). Node-RADS subcategory size also reached statistically significant difference between N0 and N1-3 stage (mean 1±0 score for N0 versus 1.3±0.47 score for N1-3, p=0.002).
ROC curve analysis for discrimination of N0 versus N1-3 with total Node-RADS-Score. The resulting area under the curve (AUC) is 0.86. A threshold value of Node-RADS 2 resulted in a sensitivity of 0.78 and a specificity of 0.86.
Also, for both readers, the Node-RADS-subcategories texture (p<0.001) and border (p=0.001) achieved statistically significance with higher subcategory-scores correlating with higher likelihood of positive N-stages. However, the subcategory shape did not achieve statistical significance (p=0.524).
ROC curve analysis for lymph node discrimination (N0 versus N1-3) using the short axis diameter resulted in an AUC of 0.87. Using a threshold value of 7.5 mm, sensitivity reached 0.78 and specificity 0.72 (Figure 3). No significant difference was found between the AUC for lymph nodes discrimination using total Node-RADS score and the AUC for discrimination with short axis diameter (p=0.85).
ROC curve analysis for discrimination of N0 vs. N1-3 with short axis diameter. Using a threshold value of 7.5 mm, sensitivity reached 0.78 and specificity 0.72.
Discrimination analysis of texture parameters for N stage. A total of 59 CT texture parameters showed statistically significant differences between the two groups (Table III). For instance, Variance (138.39±75.44 for N0 vs. 175.80±66.71 for N1-3, p=0.012, as shown in Figure 4), S(0,1)SumVarnc (398.48±27.90 vs. 419.43±19.97, p=0.002), S(0,2)SumOfSqs (101.02±8.45 vs. 107.53±5.66, p=0.011), 45dgr_GLevNonU (19.62±11.12 vs. 29.58±14.70, p=0.008). S(0,1)SumVarnc was further investigated with ROC analysis. For S(0,1)SumVarnc, an AUC of 0.75 (95%CI=0.61-0.89) was identified. A threshold value of 410.50 resulted in a sensitivity of 0.91 and a specificity of 0.67. Figure 5 displays the corresponding graph.
Texture features showing statistically significant differences between lymph nodes N0 versus lymph nodes N1-3.
Box plot of the texture feature “S(0,1)SumVarnc”. There was a statistically significant difference between nodal negative and nodal positive lymph nodes (p=0.002).
ROC analysis of the texture feature “S(0,1)SumVarnc”. It achieved an AUC of 0.75 resulting in a sensitivity of 0.91 and a specificity of 0.67.
Correlation analysis between texture features and total Node-RADS score. In correlation analysis, the following correlating texture parameters were associated with Node-RADS score with spearman correlation coefficients of r≥0.5: S(5,5)Entropy (r=0.51, p<0.001), Horzl_RLNonUni (r=0.55, p<0.001), Horzl_GLevNonU (r=0.56, p<0.001), Vertl_RLNonUni (r=0.52, p<0.001), Vertl_GLevNonU (r=0.54, p<0.001), 45dgr_RLNonUni (r=0.56, p<0.001), 45dgr_GLevNonU (r=0.56, p<0.001), 135dr_RLNonUni (r=0.54, p<0.001), 135dr_GLevNonU (r=0.55, p<0.001), WavEnLL_s-2 (r=0.53, p<0.001) and WavEnLL_s-3 (r=0.55, p<0.001).
Correlation analysis between texture features and Node-RADS subcategories.
Size. Subcategory size showed statistically significant correlations with spearman correlation coefficients of r≥0.5 with the following parameters: Horzl_RLNonUni (r=0.64, p<0.001), Horzl_GLevNonU (r=0.64, p<0.001), Vertl_RLNonUni (r=0.60, p<0.001), Vertl_GLevNonU (r=0.61, p<0.001), 45dgr_RLNonUni (r=0.64, p<0.001), 45dgr_GLevNonU (r=0.63, p<0.001), 135dr_RLNonUni (r=0.62, p<0.001), 135dr_GLevNonU (r=0.62, p<0.001).
Texture. Subcategory texture showed statistically significant correlations with spearman correlation coefficients of r≥0.35 with the following parameters: Variance (r=0.43, p<0.002), S(4,4)DifEntrp (r=0.39, p=0.005), Horzl_RLNonUni (r=0.39, p=0.005), Horzl_GLevNonU (r=0.39, p=0.006), Vertl_RLNonUni (r=0.39, p=0.005), Vertl_GLevNonU (r=0.39, p=0.006), 45dgr_RLNonUni (r=0.04, p=0.004), 45dgr_GLevNonU (r=0.4, p=0.004), 135dr_RLNonUni (r=0.37, p=0.008), 135dr_GLevNonU (r=0.38, p=0.007).
Border. Border subcategory showed statistically significant correlations with spearman correlation coefficients of r≥0.35 with the following texture features: S(0,2)Correlat (r=0.37, p=0.008), S(0,2)InvDfMom (r=0.35, p=0.013), S(0,2) SumAverg (r=−0.37, p=0.009), S(2,2)SumAverg (r=−0.36, p=0.01), S(2,-2)SumAverg (r=−0.36, p=0.01), S(0,3)Correlat (r=0.37, p=0.008), S(0,3)SumAverg (r=−0.39, p=0.006), S(0,3)SumVarnc (r=0.36, p=0.010), S(3,3)SumAverg (r=− 0.35, p=0.012), S(0,4)Correlat (r=0.36, p=0.011), S(0,4)SumAverg (r=−0.37, p=0.007), S(0,4)SumVarnc (r=0.36, p=0.011), WavEnLL_s-1 (r=0.38, p=0.006), WavEnLH_s-1 (r=−0.37, p=0.008), WavEnLL_s-2 (r=0.45, p=0.001), WavEnHL_s-2 (r=−0.38, p=0.006), WavEnLL_s-3 (r=0.46, p=0.001).
Shape. Shape subcategory showed no statistically significant correlations with spearman correlation coefficients of r≥0.35 with texture features.
Discussion
The present analysis utilized CT texture analysis as a quantitative assessment and Node-RADS as a semiquantitative method to score lymph nodes in pCCA. As key findings both imaging assessments were independently associated with the presence of lymph node metastasis and with each other.
Nodal status is of great prognostic relevance in patients with pCCA in an independent manner (28). Noteworthy, nodal positivity can be identified in 45.6% of all patients with pCCA (28). Both aspects lead to the crucial point that imaging modalities need to better diagnose malignant lymph nodes in a non-invasive manner. It could alter preoperative decision making and could deem some patients as unresectable.
The contemporary AJCC staging system considers resectable lymph node disease with nodal positive lymph nodes along the cystic duct, common bile duct, proper hepatic artery, and portal vein, whereas nodal metastasis beyond the hepatoduodenal ligament is considered as unresectable disease, which includes lymph nodes periaortic, pericaval, superior mesenteric, or celiac artery (7, 29).
Conventional CT imaging has only a limited accuracy for nodal staging. Only a pooled sensitivity of 61% and specificity of 88% in detecting nodal metastases was reported in a meta-analysis (7, 8). Moreover 13% of nodes smaller than 10 mm in short axis were in fact metastatic, and only 25% of nodes larger than 10 mm in short axis harbored disease (7). Even FDG-PET yielded a slightly higher accuracy, which highlights the need for further imaging analysis procedures to facilitate the diagnostic accuracy (30).
CT texture analysis is an emergent imaging analysis with potential diagnostic and prognostic benefits in several disease entities (10-12). It becomes clear that several texture features derived from CT and MRI images reflect distinctive histopathological tumor characteristics on a microstructure level in several tumor entities (31, 32). In a recent study texture analysis was employed to discriminate benign and malignant lymph nodes in lung cancer patients with promising results (21). It appears logical that a malignant lymph node has a different texture caused by tumor cell infiltrates compared to a benign lymph node.
Presumably, the higher heterogeneity of the lymph nodes due to tumor deposits results in a higher CT heterogeneity quantified by entropy related second order statistics, in particular. Especially for mediastinal lymph nodes, texture analysis was employed with promising results (13, 14, 21). Exemplarily in the study by Pham et al., a promising AUC of 0.89 were found in 133 malignant and 138 benign lymph nodes (33). The results of the present study suggest that texture analysis might be employed in clinical routine to better stratify patients with pCCA. Another aim of the present study was to validate the novel Node-RADS classification for the first time in patients with primary hepatic cancer.
An important field in current radiology is to standardize the radiological reporting. Node-RADS is a recently proposed scoring system to categorize several imaging findings of lymph nodes on cross sectional imaging and to aid to better report lymph node involvement in clinical routine (16). There is no restriction of this classification regarding the localization of the lymph node or the oncological disease. Node-RADS of 1 cm is proposed as general threshold value in the short axis. The present analysis revealed that Node-RADS can discriminate nodal negative from positive lymph nodes in patients with pCCA. However, it has a relatively high malignancy rate even in the lower groups 1 and 2, which could limit the clinical benefit.
An important finding of the present study is that the proposed classification has only a weak inter-reader agreement, which could further limit the possibility for translation into clinical routine. This finding can be seen in agreement with the other proposed radiological classifications (34-36). For BI-RADS, there was only a moderate agreement for mammographic criteria, even between experienced radiologists (34). Only one other study investigated the node-RADS classification in mediastinal lymph nodes also with a moderate interreader agreement (21).
This analysis provided a high malignancy rate, even in score group 1 and 2 (42.8%), which should be further discussed. A clear threshold for malignancy cannot be provided by Node-RADS, which is also in good agreement with the previous analysis of this classification on mediastinal lymph nodes (21). The malignancy rate for the Node-RADS groups 4 and 5 in the present analysis was 100% for both groups, which indicates that there might be no differences between the two categories and has no discriminative ability. Clearly, Node-RADS needs to be further elucidated in a larger sample size in terms of malignancy frequency.
Study limitations. First, it has a retrospective study design with possible known inherent bias. To reduce possible bias texture analysis and Node-RADS scoring were analyzed in a manner blinded to the clinical and pathological results. Second, the patient sample size is small. Third, CT texture analysis still lacks standardization. For potential clinical translation there is a definite need to employ the texture features investigated in other patient cohorts scanned with different CT scanners for external validation of the present results.
Conclusion
Several CT texture features and the Node-RADS score derived from staging CT were associated with the malignancy of the hilar hepatic lymph nodes and might therefore be helpful for staging. This could aid surgical treatment planning in perihilar cholangiocarcinoma.
Footnotes
Authors’ Contributions
Conception: HJM; study design: HJM and JL data collection: AS, JL, BS, AKH and RS; data analysis: HJ and JL; data interpretation: DS, TD, HJM and JL, manuscript writing: JL and HJM; manuscript editing: All Authors. The Authors read and approved the final manuscript.
Conflicts of Interest
The Authors have no conflicts of interest to declare in relation to this study.
- Received July 30, 2023.
- Revision received October 9, 2023.
- Accepted October 11, 2023.
- Copyright © 2023 International Institute of Anticancer Research (Dr. George J. Delinasios), All rights reserved.
This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY-NC-ND) 4.0 international license (https://creativecommons.org/licenses/by-nc-nd/4.0).
References
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.