The value of computed tomography-based radiomics for predicting malignant pleural effusions

Abstract

Background: Malignant pleural effusion (MPE) is a common clinical problem that requires cytological and/or histological confirmation obtained by invasive examination to establish a definitive diagnosis. Radiomics is rapidly evolving and can provide a non-invasive tool to identify MPE.

Objectives: We aimed to develop a model based on radiomic features extracted from unenhanced chest computed tomography (CT) images and investigate its value in predicting MPE.

Method: This retrospective study included patients with pleural effusions between January 2016 and June 2020. All patients underwent a chest CT scanning and medical thoracoscopy after artificial pneumothorax. Cases were divided into a training cohort and a test cohort for modelling and verifying respectively. The Mann-Whitney U test and the least absolute shrinkage and selection operator (LASSO) were applied to determine the optimal features. We built a radiomics model based on support vector machines (SVM) and evaluated its performance using ROC and calibration curve analysis.

Results: Twenty-nine patients with MPE and fifty-two patients with non-MPE were enrolled. A total of 944 radiomic features were quantitatively extracted from each sample and reduced to 14 features for modeling after selection. The AUC of the radiomics model was 0.96 (95% CI: 0.912-0.999) and 0.86 (95% CI: 0.657~1.000) in the training and test cohorts, respectively. The calibration curves for model were in good agreement between predicted and actual data.

Conclusions: The radiomics model based on unenhanced chest CT has good performance for predicting MPE and may provide a powerful tool for doctors in clinical decision-making.

Keywords: cancer; machine learning; pleural effusion; radiomics; x-ray computed tomography.

Figure 1

Flow diagram of radiomics analysis in these analyses. Core steps of radiomics analysis.

Figure 2

ROI segmentation. (A), A 62-year-old male with PE, diagnosed as tuberculous pleurisy. (B), A 67-year-old male with PE, diagnosed as malignant pleural mesothelioma. 1, Images of pleural lesions under medical thoracoscopy. 2, CT images of patients after artificial pneumothorax. 3, Regions of interest segmented on the CT under lung window. 4, The regions of interest obtained after segmentation.

Figure 3

Fourteen radiomics features selected. (A), Adjusting the parameter λ to minimize the binomial deviation of the model fitting loss value, in order to select the best radiomics characteristics. (B), The distribution of LASSO coefficients of radiomics. (C), The weights of features contributed in the model built. (D), The correlation heatmap of the fourteen features selected.

Figure 4

The receiver operator characteristic (ROC) curves of the radiomics model in the training cohort (A) and test cohort (B). AUC, area under the receiver operator characteristic curve. .

Figure 5

Calibration curves of the radiomics model in the training cohort (A) and test cohort (B). The y axis represents the actual event probability, the x axis represents the predicted event probability. The 45° dotted line represents the perfect prediction of an ideal model and the dotted lines represents the performance of the radiomics model, a closer fit to the dotted line represents a better prediction. The calibration curves indicate good calibration of the model in the training and test cohorts.