Quantitative analysis of chest MRI images for benign malignant diagnosis of pulmonary solid nodules

Abstract

Background: In this study, we developed and validated machine learning (ML) models by combining radiomic features extracted from magnetic resonance imaging (MRI) with clinicopathological factors to assess pulmonary nodule classification for benign malignant diagnosis.

Methods: A total of 333 consecutive patients with pulmonary nodules (233 in the training cohort and 100 in the validation cohort) were enrolled. A total of 2,824 radiomic features were extracted from the MRI images (CE T1w and T2w). Logistic regression (LR), Naïve Bayes (NB), support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost) classifiers were used to build the predictive models, and a radiomics score (Rad-score) was obtained for each patient after applying the best prediction model. Clinical factors and Rad-scores were used jointly to build a nomogram model based on multivariate logistic regression analysis, and the diagnostic performance of the five prediction models was evaluated using the area under the receiver operating characteristic curve (AUC).

Results: A total of 161 women (48.35%) and 172 men (51.65%) with pulmonary nodules were enrolled. Six important features were selected from the 2,145 radiomic features extracted from CE T1w and T2w images. The XGBoost classifier model achieved the highest discrimination performance with AUCs of 0.901, 0.906, and 0.851 in the training, validation, and test cohorts, respectively. The nomogram model improved the performance with AUC values of 0.918, 0.912, and 0.877 in the training, validation, and test cohorts, respectively.

Conclusion: MRI radiomic ML models demonstrated good nodule classification performance with XGBoost, which was superior to that of the other four models. The nomogram model achieved higher performance with the addition of clinical information.

Keywords: differential diagnosis; machine learning; magnetic resonance imaging; pulmonary nodules; radiomics.

Figure 1

The patient stowage flowchart was divided into training and validation sets in a ratio of 2:1.

Figure 2

Flowchart of the experimental steps. An experienced radiologist segmented the regions of interest (ROI) of the lesions. Features were selected to build the models. Receiver operating characteristic (ROC) curves were used to demonstrate the diagnostic efficiency of the models. Decision curves were used to evaluate the potential net clinical benefit of the Prediction Models.

Figure 3

Heatmap of Spearman correlations between two of the six features that were retained. The results showed that the correlation coefficients between the two pairs of features were less than 0.9.

Figure 4

(A–F) Violin plot of data distribution of the remaining six features in the two groups of benign and malignant nodules after feature selection. * <0.05, ** <0.01, *** <0.001.

Figure 5

(A–C) Diagnostic performance of machine learning models for classifying malignant and benign pulmonary nodules based on MR radiomic features in the training, validation, and test cohorts. The number adjacent to each machine learning model is the area under the receiver operating characteristic curve. LR, logistic regression; NB, Naïve Bayes; SVM, support vector machine; RF, random forest; XGBoost, extreme gradient boosting; Clinical, clinical model; COMB, combined model.

Figure 6

The radiomics nomogram was developed using independent predictors (rad score, age, and LNM) to predict benign and malignant pulmonary nodules.

Figure 7

Decision curves of the seven models in the training cohort. Net income is shown on the y-axis, and the probability threshold is shown on the x-axis.