Yuquan Zhang1,2, Chao Cheng3, Zhaobang Liu1,2, Lei Wang4, Guixia Pan3, Gaofeng Sun3, Yan Chang2, Changjing Zuo3, Xiaodong Yang2. 1. University of Science and Technology of China, Hefei, 230026, China. 2. Department of Medical Imaging, Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou, 215163, China. 3. Department of Nuclear Medicine, Changhai Hospital, Second Military Medical University, Shanghai, 200433, China. 4. School of Ophthalmology & Optometry, Eye Hospital, Wenzhou Medical University, Wenzhou, China.
Abstract
PURPOSE: To perform a radiomics analysis with comparisons of multidomain features and a variety of feature selection strategies and classifiers, with the goal of evaluating the value of quantified radiomics method for noninvasively differentiating autoimmune pancreatitis (AIP) from pancreatic ductal adenocarcinoma (PDAC) in 18 F-fluorodeoxglucose positron emission tomography/computed tomography (18 F FDG PET/CT) images. METHODS: We extracted 251 expert-designed features from 2D and 3D PET/CT images of 111 patients, and recombined these features into five feature sets according to their modalities and dimensions. Among the five feature sets, the optimal one was found leveraging four feature selection strategies and four machine learning classifiers based on the area under the receiver operating characteristic curve (AUC). The feature selection strategies include spearman's rank correlation coefficient, minimum redundancy maximum relevance, support vector machine recursive feature elimination (SVM-RFE), and no feature selection, while the classifiers are random forest, adaptive boosting, support vector machine (SVM) with the Gaussian radial basis function, and SVM with the linear kernel function respectively. Based on the optimal feature set, these feature selection strategies and classifiers were comparatively studied to achieve the best differentiation. Finally, the quantified radiomics prediction model was developed based on the best combination of the feature selection strategy and classifier, and it was compared with two clinical factors based prediction models, and human doctors using nested cross-validation in terms of AUC, accuracy, sensitivity, and specificity. RESULTS: Comparison experiments demonstrated that CT features and three-dimensional (3D) features performed better than positron emission tomography (PET) features and three-dimensional (2D) features respectively, and multidomain features were superior to single domain features. In addition, the combination of SVM-RFE feature selection strategy and Linear SVM classifier had the highest diagnostic performance (i.e., AUC = 0.93 ± 0.01, ACC = 0.85 ± 0.02, SEN = 0.86 ± 0.03, SPE = 0.84 ± 0.03). The quantified radiomics model developed is significantly superior to both human doctors and clinical factors based prediction models in terms of accuracy and specificity. CONCLUSIONS: Our preliminary results confirmed that the quantified radiomics method could aid the noninvasive differentiation of AIP and PDAC in 18 F FDG PET/CT images and the integration of multidomain features is beneficial for the differentiation.
PURPOSE: To perform a radiomics analysis with comparisons of multidomain features and a variety of feature selection strategies and classifiers, with the goal of evaluating the value of quantified radiomics method for noninvasively differentiating autoimmune pancreatitis (AIP) from pancreatic ductal adenocarcinoma (PDAC) in 18 F-fluorodeoxglucose positron emission tomography/computed tomography (18 F FDG PET/CT) images. METHODS: We extracted 251 expert-designed features from 2D and 3D PET/CT images of 111 patients, and recombined these features into five feature sets according to their modalities and dimensions. Among the five feature sets, the optimal one was found leveraging four feature selection strategies and four machine learning classifiers based on the area under the receiver operating characteristic curve (AUC). The feature selection strategies include spearman's rank correlation coefficient, minimum redundancy maximum relevance, support vector machine recursive feature elimination (SVM-RFE), and no feature selection, while the classifiers are random forest, adaptive boosting, support vector machine (SVM) with the Gaussian radial basis function, and SVM with the linear kernel function respectively. Based on the optimal feature set, these feature selection strategies and classifiers were comparatively studied to achieve the best differentiation. Finally, the quantified radiomics prediction model was developed based on the best combination of the feature selection strategy and classifier, and it was compared with two clinical factors based prediction models, and human doctors using nested cross-validation in terms of AUC, accuracy, sensitivity, and specificity. RESULTS: Comparison experiments demonstrated that CT features and three-dimensional (3D) features performed better than positron emission tomography (PET) features and three-dimensional (2D) features respectively, and multidomain features were superior to single domain features. In addition, the combination of SVM-RFE feature selection strategy and Linear SVM classifier had the highest diagnostic performance (i.e., AUC = 0.93 ± 0.01, ACC = 0.85 ± 0.02, SEN = 0.86 ± 0.03, SPE = 0.84 ± 0.03). The quantified radiomics model developed is significantly superior to both human doctors and clinical factors based prediction models in terms of accuracy and specificity. CONCLUSIONS: Our preliminary results confirmed that the quantified radiomics method could aid the noninvasive differentiation of AIP and PDAC in 18 F FDG PET/CT images and the integration of multidomain features is beneficial for the differentiation.