MRI radiomics-based machine learning model integrated with clinic-radiological features for preoperative differentiation of sinonasal inverted papilloma and malignant sinonasal tumors

• Published in: Frontiers in oncology Vol. 12; p. 1003639
• Main Authors: Gu, Jinming, Yu, Qiang, Li, Quanjiang, Peng, Juan, Lv, Fajin, Gong, Beibei, Zhang, Xiaodi
• Format: Journal Article
• Language: English
• Published: Frontiers Media S.A 23.09.2022
• Subjects: differential diagnosis; machine learning; magnetic resonance imaging; malignant sinonasal tumor; Oncology; radiomics; sinonasal inverted papilloma
• ISSN: 2234-943X; 2234-943X
• DOI: 10.3389/fonc.2022.1003639

Objective To explore the best MRI radiomics-based machine learning model for differentiation of sinonasal inverted papilloma (SNIP) and malignant sinonasal tumor (MST), and investigate whether the combination of radiomics features and clinic–radiological features can produce a superior diagnostic performance. Methods The database of 247 patients with SNIP (n=106) or MST (n=141) were analyzed. Dataset from scanner A were randomly divided into training set (n=135) and test set 1 (n=58) in a ratio of 7:3, and dataset from scanner B and C were used as an additional independent test set 2 (n=54). Fourteen clinic-radiological features were analyzed by using univariate analysis, and those with significant differences were applied to construct clinical model. Based on the radiomics features extracted from single sequence (T2WI or CE-T1WI) and combined sequence, four commonly used classifiers (logistic regression (LR), support vector machine (SVM), decision tree (DT) and k-nearest neighbor (KNN)) were employed to constitute twelve different machine learning models, and the best-performing one was confirmed as the optimal radiomics model. Furthermore, a combined model incorporated best radiomics feature subsets and clinic-radiological features was developed. The diagnostic performances of these models were assessed by the area under the receiver operating characteristic (ROC) curve (AUC) and the calibration curves. Results Five clinic-radiological features (age, convoluted cerebriform pattern sign, heterogeneity, adjacent bone involvement and infiltration of surrounding tissue) were considered to be significantly different between the tumor groups ( P < 0.05). Among the twelve machine learning models, the T2WI-SVM model exhibited optimal predictive efficacy for classification tasks on the two test sets, with the AUC of 0.878 and 0.914, respectively. For three types of diagnostic models, the combined model achieved highest AUC of 0.912 (95%CI: 0.807-0.970) and 0.927 (95%CI: 0.823-0.980) for differentiation of SNIP and MST in test 1 and test 2 sets, which performed prominently better than clinical model ( P =0.011, 0.005), but not significantly different from the optimal radiomics model ( P =0.100, 0.452). Conclusion The machine learning model based on T2WI sequence and SVM classifier achieved best performance in differentiation of SNIP and MST, and the combination of radiomics features and clinic-radiological features significantly improved the diagnostic capability of the model.