A CT-Based Radiomic Signature for the Differentiation of Pulmonary Hamartomas from Carcinoid Tumors

Abstract

Radiomics is a new image processing technology developed in recent years. In this study, CT radiomic features are evaluated to differentiate pulmonary hamartomas (PHs) from pulmonary carcinoid tumors (PCTs). A total of 138 patients (78 PCTs and 60 PHs) were evaluated. The Radcloud platform (Huiying Medical Technology Co., Ltd., Beijing, China) was used for managing the data, clinical data, and subsequent radiomics analysis. Two hand-crafted radiomics models are prepared in this study: the first model includes the data regarding all of the patients to differentiate between the groups; the second model includes 78 PCTs and 38 PHs without signs of fat tissue. The separation of the training and validation datasets was performed randomly using an (8:2) ratio and 620 random seeds. The results revealed that the MLP method (RF) was best for PH (AUC = 0.999) and PCT (AUC = 0.999) for the first model (AUC = 0.836), and PC (AUC = 0.836) in the test set for the second model. Radiomics tumor features derived from CT images are useful to differentiate the carcinoid tumors from hamartomas with high accuracy. Radiomics features may be used to differentiate PHs from PCTs with high levels of accuracy, even without the presence of fat on the CT. Advances in knowledge: CT-based radiomic holds great promise for a more accurate preoperative diagnosis of solitary pulmonary nodules (SPNs).

Keywords: carcinoid; machine learning; pulmonary hamartomas; radiomics.

Figure 1

3D segmentation, feature selection, and radiomics analysis of workflow used for PCTs and PHs.

Figure 2

The use of K best on the feature selection to further select radiomics features, which results in 11 features.

Figure 3

Lasso algorithm on feature selection. (a) Lasso path; (b) MSE path; and (c) coefficients in Lass model. Using the Lasso model, eight optimal features that correspond to the optimal alpha value were selected.

Figure 4

The figure shows the results of the 1st modeling ROC curve analysis for the training and test data for differentiating between PHs and PCTs. Note that green—PHs and red—PCTs. (a) ROC curve of the training set and (b) ROC curve test set.

Figure 5

The figure shows the results of the 2nd modeling ROC curve analysis results for the training and test data for differentiating between PHs and PCTs. Note that green—PHs and red—PCTs. (a) ROC curve of the training set and (b) ROC curve of the test set.