CT-based radiomics signature for differentiation between cardiac tumors and thrombi: a retrospective, multicenter study

The study aimed to develop and validate whether the computed tomography (CT) radiomics analysis is effective in differentiating cardiac tumors and thrombi. For this retrospective study, a radiomics model was developed on the basis of a training dataset of 192 patients (61.9 ± 13.3 years, 90 men) with cardiac masses detected in cardiac CT from January 2010 to September 2019. We constructed three models for discriminating between a cardiac tumor and a thrombus: a radiomics model, a clinical model, which included clinical and conventional CT variables, and a model that combined clinical and radiomics models. In the training dataset, the radiomics model and the combined model yielded significantly higher differentiation performance between cardiac tumors and cardiac thrombi than the clinical model (AUC 0.973 vs 0.870, p < 0.001 and AUC 0.983 vs 0.870, p < 0.001, respectively). In the external validation dataset with 63 patients (59.8 ± 13.2 years, 26 men), the combined model yielded a larger AUC compared to the clinical model (AUC 0.911 vs 0.802, p = 0.037). CT radiomics analysis is effective in differentiating cardiac tumors and thrombi. In conclusion, the combination of clinical, conventional CT, and radiomics features demonstrated an additional benefit in differentiating between cardiac tumor and thrombi compared to clinical data and conventional CT features alone.


Scientific Reports
| (2022) 12:8173 | https://doi.org/10.1038/s41598-022-12229-x www.nature.com/scientificreports/ CMR introduced the diagnostic potential of the radiomics score derived from native T1 maps to differentiate cardiac tumors from thrombi 18 . This study aimed to develop a radiomics signature and validate whether a CT radiomics analysis is effective in differentiating cardiac tumors and thrombi. In addition, we evaluated whether the CT radiomics analysis could provide additional diagnostic performance to differentiate between a cardiac tumor and a thrombus compared to clinical and conventional CT features.

Ethical considerations. This retrospective multicenter study was approved by the Institutional Review
Board of each center (Severance Hospital clinical trial center, Pusan National University Hospital clinical trial center and Gangnam Severance Hospital clinical trial center), and the requirement of obtaining informed consent was waived (Severance Hospital clinical trial center, Pusan National University Hospital clinical trial center and Gangnam Severance Hospital clinical trial center). All methods were performed in accordance with the relevant guidelines and regulations.
Study design and patient selection. In this multicenter study, a radiomic analysis was applied retrospectively to two independent data sets. For the training dataset, we retrospectively enrolled 257 consecutive patients who met the following inclusion criteria at each institution I: (1) ≥ 19 years old, (2) underwent a cardiac CT scan for a suspected cardiac mass from January 2010 to September 2019, (3) cardiac mass ≥ 1 cm on cardiac CT, and (4) underwent surgery, biopsy, or follow-up CT/transesophageal echocardiography for diagnosis of a cardiac mass. Among them, we excluded patients with unsatisfactory CT image quality (n = 35), patients whose final diagnosis was inconclusive (n = 18), or patients whose medical records such as risk factors, physical examination, and blood tests were missing (n = 12). Finally, a total of 192 patients (mean age of 61.9 ± 13.3 years, 90 men) with cardiac masses were included in the training dataset. For the external validation dataset, we retrospectively included 63 patients (mean age of 59.8 ± 13.2 years, 26 men) with cardiac masses who underwent cardiac CT from two independent institutes (Institution II and Institution III) with the same inclusion and exclusion criteria. (Fig. 1).
All data entry, data management, and analyses were coordinated or performed at a data-coordinating center. In our database, the following clinical and radiological characteristics were recorded: age, sex, history of cerebrovascular accident (CVA), atrial fibrillation or flutter, diabetes mellitus, hypertension, dyslipidemia and smoking, history of cardiac disease (including valvular heart disease, and congestive heart failure), histology subtypes, cardiac mass location, mass size, and mass density.
Cardiac computed tomography (CT) examinations. All patients were required to be fasting for at least 4 h and abstain from caffeine at least 12 h before their CT examination. Beta blockers were administered in patients with high heart rates. All cardiac CT examinations were performed using a (at minimum) 64-slice multi-detector single-or dual-source CT scanner. The type of scanner used differed among institutions and included both single-source and dual-source scanners. For the training dataset, imaging was performed using the following CT scanners: SOMATOM Definition Flash, (Siemens, GmbH, Erlangen, Germany) and Revolution CT, (GE Healthcare, Waukesha, WI, USA). For the external validation dataset, imaging was performed using the following CT scanners: SOMATOM Definition, Revolution CT, and Brilliance 64 (Philips Healthcare, Cleveland, OH, USA). All CT images were acquired according to standardized scanning protocols (Supplemen-  CT radiomic feature extraction. The process of radiomics analysis is depicted in Fig. 2 CT radiomic feature selection and model construction. Prior to feature selection and model building, no data transformation or standardization was conducted. Intraclass correlation coefficients based on a twoway mixed effect, were applied to perform a reproducibility assessment of CT radiomic features. The features with correlation coefficients > 0.75 were chosen for further analysis (Supplementary Table 2). Therefore, a total 98 of 127 features were selected as candidates for a least absolute shrinkage and selection operator (LASSO) analysis. The LASSO logistic regression model was used to select the appropriate radiomics features and build a classification model 20 . Ten-fold cross validation was performed to solve the overfitting. Features were selected if the mean of the calculated area under the curve (AUC) of the receiver operating characteristic (ROC) curve was equivalent to the maximum value. In addition, multivariable logistic regression was used to generate clinical and CT variables. Only variables with p < 0.05 in the univariable analyses were added to the final multivariable models to prevent model over-fitting. We constructed three models for discriminating between a cardiac tumor and a thrombus: a radiomics model, using CT radiomics variables; a clinical model, using clinical and conventional CT variables; a combined model, using clinical, conventional CT, and radiomics variables. Feature selection was performed only on the training dataset in order to maintain independence between the training and external validation datasets. The ROC curve was plotted to assess the differentiating performance of the three models with the training and external validation datasets.
Statistical analysis. Categorical variables were compared by the chi-square or Fisher's exact test. The differences of continuous variables were analyzed through the Student's test or the Mann-Whitney U test. Demographics, CT results, and CT radiomic features were compared between cardiac tumor and thrombus groups. The best cutoff for the predicted probability of each model in the training set was based on the Youden index (when the sum of sensitivity and specificity becomes the maximum). The model performance was compared using the ROC curve of each model to identify the model with the higher predictability. Calibration curves were performed to assess the calibration of the radiomics model, accompanied with the Brier score and the Spiegelhalter z-test. To quantify the discrimination performance of the radiomics model, AUCs were calculated in training and external validation datasets. Obuchowski method and Delong method were used to compare the performance of the ROC curves 21,22 .

Results
Patient characteristics. The study included 192 (mean age of 61.9 ± 13.3 years, 90 men) and 63 patients (mean age of 59.8 ± 13.2 years, 26 men) in the training and external validation datasets, respectively. The number of patients who underwent surgery or biopsy in the training and external validation datasets were 112 and 39, respectively. The cardiac masses in the other patients (80 in the training set and 24 in the external validation set) were diagnosed as thrombi based on the patients' response to anticoagulation treatment during follow-up. In the training dataset, there were 101 cardiac tumors and 91 thrombi in the 192 patients. Among the 101 cardiac tumors, 82 were benign, including myxoma (n = 69), papillary fibroelastoma (n = 4), cavernous hemangioma (n = 3), lipoma (n = 3), and fibroma or angiofibroma (n = 3). Nineteen were malignant, including metastasis (n = 11), lymphoma (n = 7) and sarcoma (n = 1). In the external dataset, there were 38 cardiac tumors and 25 thrombi in the 63 patients. Among the 38 cardiac tumors, 36 were myxoma and 2 were sarcoma. All tumors were confirmed by surgical excision or biopsy.
The baseline characteristics of all patients are summarized in Table 1. In the training dataset, patients with thrombi had higher incidences of history of CVA (p = 0.004), atrial fibrillation or flutter (p < 0.001), and history of cardiac disease (p < 0.001) than those with cardiac tumors. Other clinical characteristics of the two groups were not significantly different. The mean diameter of cardiac tumors was significantly larger than that of thrombi (35.7 ± 16.7 vs 27.7 ± 15.0 mm, p < 0.001). In the external validation dataset, patients with thrombi had higher incidences of a history of atrial fibrillation or flutter (p = 0.005) than those with cardiac tumors. The mean diameter of the cardiac tumors was significantly larger than that of thrombi (33.9 ± 15.2 vs 26.2 ± 13.7 mm, p = 0.045).

Discussion
This study was designed to develop and validate whether the CT radiomics analysis technique is effective in differentiating cardiac tumors and thrombi. The main finding was that a combination of clinical, conventional CT, and radiomics features demonstrated additional benefits in differentiating between cardiac tumors and thrombi compared to clinical model (clinical and conventional CT features) alone.
Distinguishing a cardiac thrombus from a tumor is challenging because the clinical and radiological signs are very similar, but the subsequent medical treatment is different. CMR is currently the standard method for evaluating cardiac mass 5 . Previous studies have reported that late gadolinium enhancement-CMR can be useful for differentiating between cardiac tumors and thrombi [23][24][25] .
Cardiac CT has been proposed to help differentiate between cardiac tumors and thrombi as part of a multimodality approach. Although useful imaging findings of cardiac tumors and thrombi can be found on CT scans, the potential overlap in the imaging findings often may result in persisting uncertainty in the differentiation of the two diseases. Previous studies have used only visual assessment or HU-based CT values to differentiate between cardiac tumors and thrombi in CT [26][27][28] . The problem with these analyses is moderate inter-reader reproducibility of visual assessment and only rely on the distribution of CT numbers, which shows major overlaps between two disease entities. In our present study, a larger cardiac mass (OR, 1.09; 95% CI, 1.04-1.07; p = 0.034) was significantly associated with the prediction of a cardiac tumor, while the CT-based HU value was not predictive for either cardiac tumor or thrombus. As different disease entities require different management strategies, it is crucial to obtain a correct diagnosis in the most noninvasive way as possible.
Radiomics is an emerging field of study that allows to extract quantitative imaging features from radiological datasets and describe the heterogeneity and spatial complexity of given regions of interest. This feature generative technique makes it possible to precise identification of phenotype abnormalities in medical images and may provide additional information, potentially allowing histological classification of abnormalities based on the images 29 . In this study, we constructed three models for discriminating between cardiac tumor and cardiac thrombus and compared the model performance in testing and validation datasets. Our radiomics model showed good discriminatory performance between cardiac tumors and thrombi with AUC of 0.973 and 0.872 in training and validation datasets, respectively. When we combined the radiomics model with the clinical model, our combined model showed significantly higher discriminating performance between cardiac tumors and thrombi than that of the clinical model (clinical and conventional CT parameters) in an external validation dataset (AUC 0.911 vs 0.802, p = 0.037). In both training and external datasets, sensitivity and diagnostic accuracy of combined models were higher than those of clinical model. This result validated and supported that radiomics may have additional value in the differentiation between cardiac tumors and thrombi using CT. Recent investigation demonstrated that when histological classification of coronary lesions can be predicted using a radiomics-based machine learning model, which outperformed visual assessment 30,31 . It seems that radiomics can extract new information from medical images, which can improve the discriminatory power of current medical devices.      www.nature.com/scientificreports/ Despite the novel analysis in this study, several limitations must be acknowledged. First, the proposed radiomics model was established on the basis of data obtained from a single center. Although our radiomics model was validated with an external validation dataset, prospective multicenter studies with considerably large datasets are needed to further validate the robustness and reproducibility of our radiomics analysis. Second, the CT protocol can influence the results of the radiomics approach. In our study, image standardization was not performed when constructing the CT radiomics model. For CT radiomics, image reconstruction algorithms (i.e., reconstruction kernels) and section thickness have been major sources of radiomic feature variability 32,33 . A previous study demonstrated that different kernels significantly reduced the reproducibility of radiomic features, with only 15.2% of radiomic features were reliable when using different reconstruction kernel 33 . Although the CT images were not standardized, the results of our radiomics model improved the ability to differentiate between cardiac tumors and thrombi in external validation datasets with different CT images.
In conclusion, CT radiomics analysis is effective in differentiating cardiac tumors and thrombi. The combination of clinical, conventional CT, and radiomics features demonstrated additional benefits in differentiating between cardiac tumors and thrombi compared to clinical and conventional CT features alone. Hence, the combined model may be useful to differentiate between cardiac tumors and thrombi when other imaging modalities are inconclusive.