Ruiping Zhang1, Lei Zhu2, Zhengting Cai3, Wei Jiang2, Jian Li4, Chengwen Yang1, Chunxu Yu5, Bo Jiang1, Wei Wang1, Wengui Xu2, Xiangfei Chai3, Xiaodong Zhang6, Yong Tang7. 1. Department of Radiation Oncology, Department of Radiation Physics, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center of Cancer, Key Laboratory of Cancer Prevention and Therapy, Huanhu West Road, Hexi District, Tianjin, 300060, China. 2. Department of Nuclear Medicine, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center of Cancer, Key Laboratory of Cancer Prevention and Therapy, Huanhu West Road, Hexi District, Tianjin, 300060, China. 3. Huiying Medical Technology Co., Ltd., Dongcheng Science and Technology Park, Dongcheng Road, Haidian District, Beijing, 100192, China. 4. Department of Pancreas Cancer, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center of Cancer, Key Laboratory of Cancer Prevention and Therapy, Huanhu West Road, Hexi District, Tianjin, 300060, China. 5. Department of Physics, Nankai University, Weijin Road, Nankai District, Tianjin, 300071, China. 6. Department of Radiation Oncology, Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, 1840 Old Spanish Trail, Houston, TX, 77054, USA. 7. Department of Pancreas Cancer, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center of Cancer, Key Laboratory of Cancer Prevention and Therapy, Huanhu West Road, Hexi District, Tianjin, 300060, China. Electronic address: tli001@163.com.
Abstract
PURPOSE: The study is to explore potential features and develop classification models for distinguishing benign and malignant lung lesions based on CT-radiomics features and PET metabolic parameters extracted from PET/CT images. MATERIALS AND METHODS: A retrospective study was conducted in baseline 18 F-flurodeoxyglucose positron emission tomography/ computed tomography (18 F-FDG PET/CT) images of 135 patients. The dataset was utilized for feature extraction of CT-radiomics features and PET metabolic parameters based on volume of interest, then went through feature selection and model development with strategy of five-fold cross-validation. Specifically, model development used support vector machine, PET metabolic parameters selection used Akaike's information criterion, and CT-radiomics were reduced by the least absolute shrinkage and selection operator method then forward selection approach. The diagnostic performances of CT-radiomics, PET metabolic parameters and combination of both were illustrated by receiver operating characteristic (ROC) curves, and compared by Delong test. Five groups of selected PET metabolic parameters and CT-radiomics were counted, and potential features were found and analyzed with Mann-Whitney U test. RESULTS: The CT-radiomics, PET metabolic parameters, and combination of both among five subsets showed mean area under the curve (AUC) of 0.820 ± 0.053, 0.874 ± 0.081, and 0.887 ± 0.046, respectively. No significant differences in ROC among models were observed through pairwise comparison in each fold (P-value from 0.09 to 0.81, Delong test). The potential features were found to be SurfaceVolumeRatio and SUVpeak (P < 0.001 of both, U test). CONCLUSION: The classification models developed by CT-radiomics features and PET metabolic parameters based on PET/CT images have substantial diagnostic capacity on lung lesions.
PURPOSE: The study is to explore potential features and develop classification models for distinguishing benign and malignant lung lesions based on CT-radiomics features and PET metabolic parameters extracted from PET/CT images. MATERIALS AND METHODS: A retrospective study was conducted in baseline 18 F-flurodeoxyglucose positron emission tomography/ computed tomography (18 F-FDG PET/CT) images of 135 patients. The dataset was utilized for feature extraction of CT-radiomics features and PET metabolic parameters based on volume of interest, then went through feature selection and model development with strategy of five-fold cross-validation. Specifically, model development used support vector machine, PET metabolic parameters selection used Akaike's information criterion, and CT-radiomics were reduced by the least absolute shrinkage and selection operator method then forward selection approach. The diagnostic performances of CT-radiomics, PET metabolic parameters and combination of both were illustrated by receiver operating characteristic (ROC) curves, and compared by Delong test. Five groups of selected PET metabolic parameters and CT-radiomics were counted, and potential features were found and analyzed with Mann-Whitney U test. RESULTS: The CT-radiomics, PET metabolic parameters, and combination of both among five subsets showed mean area under the curve (AUC) of 0.820 ± 0.053, 0.874 ± 0.081, and 0.887 ± 0.046, respectively. No significant differences in ROC among models were observed through pairwise comparison in each fold (P-value from 0.09 to 0.81, Delong test). The potential features were found to be SurfaceVolumeRatio and SUVpeak (P < 0.001 of both, U test). CONCLUSION: The classification models developed by CT-radiomics features and PET metabolic parameters based on PET/CT images have substantial diagnostic capacity on lung lesions.
Authors: Marcus Makowski; Tobias Penzkofer; Boris Gorodetski; Philipp Hendrik Becker; Alexander Daniel Jacques Baur; Alexander Hartenstein; Julian Manuel Michael Rogasch; Christian Furth; Holger Amthauer; Bernd Hamm Journal: Eur Radiol Exp Date: 2022-09-15