Deep learning outperformed radiomics based on MRI in the differentiation of sinonasal small round cell and non-small round cell malignant tumors.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiology, Beijing TongRen Hospital, Capital Medical University, Beijing, China.
- School of Biomedical Engineering, ShanghaiTech University, Shanghai, China; Artificial Intelligence & Neuro-Informatics Engineering (ARINE) Laboratory, School of Computer Engineering, Jiangsu Ocean University, Lianyungang, China. Electronic address: [email protected].
- Department of Radiology, Beijing TongRen Hospital, Capital Medical University, Beijing, China. Electronic address: [email protected].
Abstract
Differentiating sinonasal small round cell malignant tumors (SRCMTs) from non-SRCMTs is challenging due to overlapping MRI features. This study aimed to compare the diagnostic performance of deep learning and radiomics models for preoperative MRI-based classification. We retrospectively analyzed 325 patients with pathologically confirmed sinonasal malignancies (163 SRCMTs and 162 non-SRCMTs). Each patient underwent T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and contrast-enhanced T1-weighted imaging (CE-T1WI). Tumors were manually segmented. Five convolutional neural networks (CNNs)-ResNet-18, ResNet-34, ResNet-50, VGG13, and VGG16-were trained on each sequence. For radiomics, 1200 features were extracted per sequence, and multiple machine learning classifiers were trained. Model performance was assessed by area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). The DeLong test was used to compare AUCs between models, with statistical significance set at P < 0.05. The CE-T1WI-based ResNet-34 model achieved the best performance, with an average AUC of 0.830, the accuracy of 0.755, sensitivity of 0.918, specificity of 0.592, PPV of 0.692, and NPV of 0.879. The corresponding CE-T1WI-based radiomics model using a support vector machine yielded an AUC of 0.758 (accuracy = 0.755, sensitivity = 0.840, specificity = 0.667, PPV = 0.724, NPV = 0.800). On the independent test cohort, ResNet-34 showed numerically higher discriminative performance than the radiomics model, although this difference did not reach statistical significance. For T1WI and T2WI, deep learning and radiomics models demonstrated broadly comparable performance. A CE-T1WI-based ResNet-34 network provided high diagnostic efficacy, and in our cohort deep learning models achieved numerically higher comparable performance to MRI-based radiomics models for differentiating SRCMTs from non-SRCMTs.