To develop a multimodal MRI-based radiomics model for the differential diagnosis of benign and malignant lung lesions, and to compare the discriminative abilities of different models.

Totally 114 patients with 115 lesions (44 benign and 71 malignant) in Nantong First Peoples′s Hospital from January 2014 to October 2019 were included in the study. All patients underwent non-enhanced MR examination, and textural features from T₁WI,T₂WI and apparent diffusion coefficient (ADC) imaging were extracted. The feature selection methods included L1 based, mutual information, tree based, recursive feature elimination and F-test. Then we constructed a prediction model by using logistic regression (LR), support vector machine (SVM), random forest (RF) and k-nearest neighbor (KNN) respectively. In order to control the number of modeling features and reduce the ininterpretability of the model, the new model was obtained by manually modifying some parameters of the hyperparameter model. One hundred and fourteen cases were rotated as training and validation sets. The performance of each model was evaluated by confounding matrix and receiver operating characteristic (ROC) curve.

The area under the curve (AUC) of T₂WI based LR model for the differential diagnosis of benign and malignant pulmonary nodules/masses was 0.71 and the F1 score was 0.57. Based on T₁WI images, LR and SVM model could be used to identify benign and malignant pulmonary nodules, the AUC before parameter adjustment were 0.77 and 0.78, the accuracy after parameter adjustment (LR^a,SVM^a) was 0.67, 0.70, and both the AUC were 0.72. However, no matter which feature or classifier was selected, both the AUC and accuracy of ADC-based model were less than 0.70.

Multimodal MRI-based radiomics model is valuable for the differential diagnosis of benign and malignant pulmonary nodules/masses, and T₁WI-based model shows the best discrimination.

Lung neoplasms; Magnetic resonance imaging; Diagnosis, differential; Radiomics