Advanced biomedical imaging for accurate discrimination and prognostication of mediastinal masses

To investigate the potential of radiomic features and dual‐source dual‐energy CT (DECT) parameters in differentiating between benign and malignant mediastinal masses and predicting patient outcomes.


| INTRODUCTION
The mediastinum is a crucial anatomical region that contains vital structures including the heart, great vessels, trachea and oesophagus. 1 Mediastinal masses constitute of a broad spectrum of entities that can be either benign or malignant. 2,3Thymic epithelial tumours, which include thymomas and thymic malignancies, account for nearly half of all anterior mediastinal tumours. 46][7] Differentiating between benign and malignant mediastinal masses is essential for appropriate clinical management and treatment planning, as treatment approaches and prognoses differ significantly. 8Accurate diagnosis relies on the integration of clinical and radiological findings, emphasizing the importance of radiologists' familiarity with the spectrum of imaging findings on CT and MRI when evaluating mediastinal masses. 9dvancements in cancer imaging have brought about significant progress in the development of imaging biomarkers.Dual-energy CT (DECT) has emerged as a valuable tool, allowing for material decomposition by analysing energy attenuation behaviour of tissues across different energy levels. 10,11This technique facilitates the development of material decomposition methods like fat fraction analysis and iodine mapping, providing additional insights into tissue composition. 12Radiomics is a noninvasive approach that enables the extraction of high-dimensional image data, going beyond visual interpretation to gain a more comprehensive understanding of tissue morphology. 13,14It is a promising tool in the field of oncology as it has the ability to aid in the noninvasive classification of tumour subtypes and to assess treatment response. 15he current classification of mediastinal masses as benign or malignant relies primarily on subjective assessment and invasive biopsy procedures. 16,17To address this limitation, the objective of this study was to explore a multiparametric approach utilizing radiomic features and DECT-derived imaging biomarkers.The aim was to noninvasively differentiate between benign and malignant mediastinal masses and predict patient outcomes.

| Institutional review board approval
The present retrospective single-centre study received approval from the institutional ethical review board, which granted a waiver for written informed consent.

| Study population
2014 and January 2023 were retrospectively identified and extracted from our institutional CT database.To maintain the accuracy of the study, scans affected by motion artefacts (n = 2) and cases with missing reference standard (n = 2) were excluded from the initial study population.The final study cohort consisted of 90 patients.The study inclusion process is illustrated in Figure 1.
Histopathological confirmation served as the reference standard for all mediastinal masses included in this study.The mediastinal masses were classified into two categories: benign (including thymic hyperplasia/rebound [n = 10], mediastinitis [n = 16] and thymoma [n = 23]) and malignant (including lymphoma [n = 28], mediastinal tumour [n = 4] and thymic carcinoma [n = 9]).Clinical data, including date of birth, gender, tumour stage, laboratory parameters and follow-up information, were extracted from all available electronic medical records obtained during routine clinical practice.Table 1 provides an overview of the patient characteristics.

| DECT scan protocol
The DECT examinations were conducted using a thirdgeneration dual-source DECT scanner (Somatom Force, Siemens Healthineers).The two x-ray tubes were adjusted to different kV tube voltages (tube A: 90 kVp and 145 mAs; tube B: Sn150 kVp [.64-mm tin filter] and 90 mAs).The scanning of patients was performed in craniocaudal direction during inspiratory breath-hold.The scanning parameters included a rotation time of .5 s, collimation of 2 × 192 × .6 mm, and a pitch of .6.Prior to scanning, a nonionic contrast agent (Imeron 350, Bracco) was administered intravenously to all patients at a flow rate of 2-3 mL/s, with a dose of 1.2 mL/kg of body weight (maximum 120 mL).The acquisition of contrast-enhanced DECT imaging data sets occurred 50 s after the injection of the contrast media.
For image reconstruction, an iterative reconstruction algorithm (ADMIRE®, Siemens Healthineers) was utilized.The reconstructed images had a section thickness of 3 mm and an increment of 2 mm.DECT post-processing was performed using a 3D multi-modality workstation (syngo.via,version VB10B, Siemens Healthineers) with an iodine subtraction package (Liver VNC, Siemens Healthineers).To facilitate radiomics analysis, the image stacks were exported in DICOM format and imported into the 3D Slicer software platform (http://slicer.org,Version 4.9.0). 4    interest (ROI) measurements for DECT material decomposition analysis, whereas one reader (V.K., radiologist in training with 3 years of experience in experimental imaging) conducted volume of interest (VOI) measurements for radiomic analysis.All DECT ROI measurements were drawn with a maximum diameter of 1 cm, excluding surrounding structures such as mediastinal fat, vessels and calcifications.For radiomic analysis, VOI measurements were performed by segmenting the entire mediastinal lesion (Figure 2).The values for attenuation, iodine density and fat fraction were determined from contrastenhanced series, while radiomic features were extracted from both contrast-enhanced and noncontrast series.To extract the radiomic features, PyRadiomics, an opensource extension integrated into the 3D Slicer software platform, was utilized.The extracted features were classified into seven categories: shape, first-order, gray-level dependence matrix (GLDM), gray-level co-occurrence matrix (GLCM), grey-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM) and neighbouring gray tone difference matrix (NGTDM).For this study, a radiomics quality score of 22 was calculated using the https://www.radiomics.world/rqs2 tool to assess the quality and reliability of the radiomics analysis.To give an overview of the transparency and quality of the study, a CheckList for EvaluAtion of Radiomics research (CLEAR) 21 is given in Table S1.

| Statistical analysis
For statistical analysis, we used MedCalc (Version 20.123).The Kolmogorov-Smirnov test was used to assess the normality of data.The analysis of variance (ANOVA) test was used utilized for data with a continuous distribution, while the Wilcoxon signed-rank test was used for nonnormally distributed data.We considered a p < .05 to be statistically significant.
To evaluate inter-rater reliability, we calculated the intra-class correlation coefficient (ICC) using a two-way mixed-effects model.The ICC values were interpreted as poor agreement (ICC < .50),moderate agreement (ICC .50-.75), good agreement (ICC .75-.90) or excellent agreement (ICC > .9). 22 Mean values of DECTderived imaging parameters and radiomic features were compared between benign and malignant mediastinal masses.The radiomic features were randomly divided into training (60%) and test (40%) datasets.Cox proportional hazards models were used to identify independent factors among DECT-derived imaging markers and radiomic features.Overall survival was defined as the time from the DECT scan until death or the last
recorded follow-up.Multivariate Cox proportional hazards models were adjusted for significant univariate prognostic parameters and clinically relevant confounders.Hazard ratios with corresponding 95% confidence intervals (CIs) were reported.The discriminative ability of the model was evaluated by plotting the receiver operating characteristic (ROC) curves and calculating the area under the curve (AUC) for the assessment of mediastinal masses and survival.

| Diagnostic performance of DECT imaging parameters and radiomic features
In the contrast-enhanced CT analysis, the overall diagnostic performance of CT radiomic features and DECT imaging parameters in discriminating between benign and mediastinal masses was excellent (AUC .98,95% CI, .89-1.00; p < .001).Summarized details with AUCs as well as positive and negative predictive values for all radiomic features and DECT imaging parameters are displayed in Table 3.
The Cox regression analysis revealed that the radiomic features, as a whole, provided independent prognostic information for death, with a c-index of .64 (95% CI, .50-.77, p = .028).After adjusting the unadjusted radiomics model by including DECT imaging parameters that showed significant associations in the univariate analysis (iodine density and attenuation), the prognostic power improved, yielding a c-index of .73(95% CI, .60-.85, p = .008).By adding DECT imaging parameters as well as clinical parameters that reached univariate significance (iodine density, attenuation and gender) to the unadjusted radiomics model, the Chi-square value significantly increased from 4.81 (c-index = .64[95% CI, .50-.77], p = .028)to 19.91 (cindex = .80(.70-.89), p < .001).Table 4 summarizes the performances of the different Cox regression models in predicting outcomes of patients with mediastinal masses, incorporating radiomic features, DECT imaging parameters and gender.

| DISCUSSION
An accurate and reliable differentiation between benign and malignant mediastinal masses is crucial for effective clinical management and treatment planning. 17,23In this study, we aimed to explore a multiparametric approach based on radiomic features and DECT-derived imaging biomarkers to noninvasively distinguish between benign and malignant mediastinal masses and predict patient outcomes.To our knowledge, this is the first study to demonstrate the utility of a combined application of DECT imaging biomarkers and radiomic features in discriminating between benign and malignant mediastinal masses (AUC .980).Our findings suggest that several DECT imaging parameters and radiomic features can contribute to the prediction of outcomes in patients with mediastinal masses.The multiparametric approach, incorporating DECT imaging parameters, radiomic features and clinical parameters showed strong prognostic power in predicting survival (model 4; c-index, .796[.702-.890],p = .005).
The emergence of dual-source DECT has provided an opportunity to characterize tissue by acquiring CT images at two different energy spectra. 24,25One particular aspect that has garnered scientific interest in recent decades is the measurement of iodine concentration in organs and pathological lesions, which allows for the generation of iodine maps. 24,26Previous studies in this field have demonstrated notable distinctions in DECT-derived imaging biomarkers when differentiating mediastinal masses.In a study from 2022, the authors investigated the impact of DECT-derived fat fraction measurements for the discrimination between mediastinal lesions and normal thymic tissue. 27In another study from 2022, the authors successfully validated the use of DECT-derived spectral parameters, in particular iodine concentration and normalized F I G U R E 3 Case example of a 26-year-old male with a germ cell tumour.This 26-year-old male presented to the emergency department with shortness of breath and fever.Chest x-ray demonstrated an enlarged superior mediastinum.The subsequently performed CT scan confirmed the finding that was reported on the chest x-ray and showed a large anterior mediastinal mass with central areas of necrosis.A subcohort (thymic pathologies) of the current study has been previously reported. 18bbreviations: DECT, dual-energy CT; HU, hounsfield units; SD, standard deviation.iodine concentration, to differentiate between thymic carcinoma and invasive thymoma in DECT perfusion imaging. 28However, these studies did not specifically explore the differences in DECT-derived imaging parameters between benign and malignant mediastinal masses.

T A B L E 3
In contrast, our study successfully confirmed variations in DECT-derived iodine density and fat fraction, demonstrating their potential in discriminating between benign and malignant mediastinal masses in general.This finding highlights the significant potential of DECT postprocessing techniques in providing additional insights into contrast enhancement, angiogenesis and mass heterogeneity.Accurate discrimination between benign and malignant mediastinal masses plays a crucial role in determining patient prognosis, and the utilization of DECTderived imaging parameters can contribute to improved clinical decision-making. 8n addition to DECT post-processing techniques, the utilization of radiomic features has provided a valuable tool for tissue characterization by extracting highdimensional image data. 28Several previous studies have highlighted the potential of radiomic features in distinguishing masses within the anterior mediastinum. 29,30or instance, a study conducted in 2022 investigated the impact of a radiomics-based model in assisting clinical decision-making for patients with anterior mediastinal masses in a preoperative setting. 29The authors successfully developed a combined model based on radiomic features and clinical parameters derived from noncontrast CT scans, which exhibited good preoperative diagnostic accuracy in predicting the need for therapeutic thymectomy (AUC .870).In another recent study, the authors developed a radiomics-based model to aid at predicting pathological diagnoses of anterior mediastinal masses, encompassing both benign lesions and malignant thymic tumours. 30The study revealed moderate diagnostic performance in differentiating benign from malignant lesions (AUC .715)and good diagnostic performance in distinguishing thymomas from thymic carcinomas (AUC .810).Building upon these findings, our multiparametric approach, which incorporates not only radiomic features and clinical parameters but also DECT-derived imaging parameters, has the potential to further improve the discrimination of anterior mediastinal masses.Our approach exhibits even higher AUC values (AUC .980) in the differentiation between benign and malignant thymic masses, surpassing previously reported results.Altogether, the findings of our studies are in line with previous studies that investigated a multiparametric approach involving radiomic features, DECT parameters and clinical parameters to discriminate between tumours of different degrees of malignancy. 18,31onsidering the broad range of anterior mediastinal lesions, our methodology offers a comprehensive framework to improve the characterization and classification of mediastinal tumours.
Interestingly, our study revealed that significant variations in radiomic features between benign and malignant mediastinal masses were also found for noncontrast DECT scans, allowing for discrimination between benign and malignant mediastinal masses with an AUC of .899.This finding holds particular significance for patients with medical conditions that require the reduction or absence of contrast media, such as acute kidney failure, chronic renal insufficiency, hyperthyroidism or contrast media allergies. 32In such cases, the extraction of radiomic features from noncontrast CT scans offers a promising alternative to standard contrast-enhanced CT scans with only slightly inferior AUCs in noncontrast versus contrast-enhanced scans (AUC .899 vs. AUC .980).This suggests that noncontrast CT scans, with their lower risk and broader applicability, can still provide valuable information for accurate and treatment planning.Our findings demonstrate the ability of radiomics analysis to capture subtle differences in tissue morphology beyond visual perception in both contrast-enhanced and non-contrast CT scans, potentially assisting in the noninvasive classification of tumour subtypes.By incorporating multiple imaging parameters and clinical data, the approach utilized in our study shows promise in improving the accuracy of noninvasive diagnosis and might reduce the need for invasive biopsy procedures in the future.
It is important to acknowledge the limitations of our study.First, the retrospective design and single-institution setting may introduce selection bias.To overcome this limitation, prospective multicentre studies with larger sample sizes are warranted to validate our findings.Second, all subtypes of thymomas were analysed together, overlooking the potential differences in invasiveness among the subtypes, particularly type B thymomas.Third, the generalizability of our results may be limited to the specific DECT scanner and imaging protocol used in our study.Future studies should investigate the reproducibility of our findings across different DECT platforms.Also, while our multiparametric approach shows promise in noninvasive classification and diagnosis, its clinical utility should be evaluated through prospective clinical trials.These trials would assess the impact of the proposed approach on treatment decision-making and patient outcomes, providing a more comprehensive understanding of its practical implications.Last, our study yielded a radiomics quality score of 22.The radiomics quality score gives an overview of the transparency of the study and facilitates the repeatability and reproducibility of radiomics research.Although comparable to previous studies, a higher radiomics quality score would have further enhanced the power of our findings.
In conclusion, our study highlights the potential of an approach that combines radiomic features and DECTderived imaging biomarkers for the noninvasive differentiation of benign and malignant mediastinal masses.The observed significant differences in DECT-derived imaging biomarkers and radiomic features provide valuable information for the characterization of these masses.Incorporating these advanced imaging techniques into clinical practice may enhance the management and treatment planning for patients with mediastinal masses, ultimately improving patient care and outcomes.Further research and validation are needed to fully establish the clinical utility of these techniques and their impact on patient management.
Two readers with different levels of radiological experience (L.D.G., radiologist in training with 3 years of experience in experimental imaging and C.B., boardcertified-radiologist with 8 years of experience in experimental imaging) independently conducted region of F I G U R E 1 STARD flowchart of study inclusion.T A B L E 1 Baseline characteristics of the study population.

F I G U R E 2
Illustration of the radiomics segmentation process.Illustration of the radiomics segmentation process in contrast-enhanced dual-energy CT on a patient with histopathologically confirmed lymphoma.(A) shows an axial image of the segmented lesion in contrastenhanced chest dual-energy CT.Threedimensional models of the segmented lesion are displayed in (B) and (C).
(A) and (B) illustrate contrast-enhanced scans of the patient in axial and coronal orientation, respectively.(C) illustrates a DECT-derived iodine perfusion image in axial orientation.A CT-guided biopsy was performed and histopathology confirmed the diagnosis of a germ cell tumour.T A B L E 2 Comparison of dual-energy CT-derived imaging biomarkers.
Performance of the imaging biomarkers in contrast-enhanced CT to discriminate between benign and malignant mediastinal masses.: AUC, the area under the curve; CT, computed tomography; GLCM, gray-level co-occurrence matrix; GLDM, gray-level dependence matrix; GLRLM, grey-level run length matrix; GLSZM, gray-level size zone matrix; HU, hounsfield unit; RU, relative unit; NGTDM, neighbouring gray tone difference matrix; NPV, negative predictive value; PPV, positive predictive value. Abbreviations

TABLE 3 (
Continued) Performance of Cox regression models to predict outcome combining radiomic features, dual-energy CT imaging parameters and clinical characteristics.
T A B L E 4