Xiao-Jie Xie1, Si-Yun Liu2, Jian-You Chen3, Yi Zhao4, Jie Jiang1, Li Wu1, Xing-Wen Zhang5, Yi Wu5, Hui Duan1, Bing He4, Heng Luo6, Dan Han7. 1. Department of Medical Imaging, First Affiliated Hospital of Kunming Medical University, Kunming, Yunnan, 650032, China. 2. Precision Health Institution, GE Healthcare (China), Beijing, 100176, China. 3. Department of Radiology, Third Affiliated Hospital of Kunming Medical University, Kunming, Yunnan, 650106, China. 4. Department of Pathology, the People's Hospital of Chuxiong Yi Autonomous Prefecture, Chuxiong, Yunnan, 675099, China. 5. Department of Radiology, the People's Hospital of Chuxiong Yi Autonomous Prefecture, Chuxiong, Yunnan, 675099, China. 6. Office of the Vice President, the People's Hospital of Chuxiong Yi Autonomous Prefecture, Chuxiong, Yunnan, 675099, China. Electronic address: lamskan25@163.com. 7. Department of Medical Imaging, First Affiliated Hospital of Kunming Medical University, Kunming, Yunnan, 650032, China. Electronic address: kmhandan@sina.com.
Abstract
OBJECTIVES: We aimed to explore the feasibility of 2D and 3D radiomics signature based on the unenhanced computed tomography (CT) images to predict BRCA1-associated protein 1 (BAP1) gene mutation status for malignant pleural mesothelioma (MPM) patients. MATERIALS AND METHODS: 74 patients with MPM were retrospectively enrolled (22 mutant BAP1, 52 wild-type BAP1 demonstrated by Sanger sequencing). The radiomic features were extracted respectively from the 2D and 3D segmentation of unenhanced pre-treatment CT images, and the dataset was randomly divided into training (n = 51) and test (n = 23) sets for radiomics model development and internal validation. The synthetic minority over-sampling technique (SMOTE) was used for data balancing in the training set. 2D or 3D features were sequentially selected by ICC > 0.8, correlation analysis (cut-value 0.7), univariate analysis or univariate logistic regression (LR), which were involved into multivariate LR for LR model construction. Following the comparison of the 2D and 3D models by the ROC analysis and Delong test for AUC, the calibration and clinical utility of 2D and 3D models were evaluated. RESULTS: 3D radiomic features showed better ICCs compared with 2D in both intra- (P < 0.001) and inter-observer (P < 0.001) analysis. 3D radiomic model based on selected features developed from a balanced training dataset presented a favorable predictive performance with AUC of 0.786 and 0.768 in the training and test sets, respectively. The predictive performance of 3D model was superior to 2D model (1 feature) both in the training (AUC 0.786 vs. 0.683, P = 0.036) and the test (AUC 0.768 vs.0.652, P = 0.441) set. The calibration curve and decision curves also indicate a better BAP1 prediction performance and clinical benefit for 3D model than that of 2D model. CONCLUSION: The developed unenhanced CT-based 3D radiomics signature is potential as a noninvasive marker for predicting BAP1 mutation status.
OBJECTIVES: We aimed to explore the feasibility of 2D and 3D radiomics signature based on the unenhanced computed tomography (CT) images to predict BRCA1-associated protein 1 (BAP1) gene mutation status for malignant pleural mesothelioma (MPM) patients. MATERIALS AND METHODS: 74 patients with MPM were retrospectively enrolled (22 mutant BAP1, 52 wild-type BAP1 demonstrated by Sanger sequencing). The radiomic features were extracted respectively from the 2D and 3D segmentation of unenhanced pre-treatment CT images, and the dataset was randomly divided into training (n = 51) and test (n = 23) sets for radiomics model development and internal validation. The synthetic minority over-sampling technique (SMOTE) was used for data balancing in the training set. 2D or 3D features were sequentially selected by ICC > 0.8, correlation analysis (cut-value 0.7), univariate analysis or univariate logistic regression (LR), which were involved into multivariate LR for LR model construction. Following the comparison of the 2D and 3D models by the ROC analysis and Delong test for AUC, the calibration and clinical utility of 2D and 3D models were evaluated. RESULTS: 3D radiomic features showed better ICCs compared with 2D in both intra- (P < 0.001) and inter-observer (P < 0.001) analysis. 3D radiomic model based on selected features developed from a balanced training dataset presented a favorable predictive performance with AUC of 0.786 and 0.768 in the training and test sets, respectively. The predictive performance of 3D model was superior to 2D model (1 feature) both in the training (AUC 0.786 vs. 0.683, P = 0.036) and the test (AUC 0.768 vs.0.652, P = 0.441) set. The calibration curve and decision curves also indicate a better BAP1 prediction performance and clinical benefit for 3D model than that of 2D model. CONCLUSION: The developed unenhanced CT-based 3D radiomics signature is potential as a noninvasive marker for predicting BAP1 mutation status.
Authors: Luigi Vimercati; Domenica Cavone; Francesco Fortarezza; Maria Celeste Delfino; Romina Ficarella; Angela Gentile; Angela De Palma; Giuseppe Marulli; Luigi De Maria; Concetta Caporusso; Andrea Marzullo; Antonio d'Amati; Daniele Egidio Romano; Antonio Caputi; Stefania Sponselli; Gabriella Serio; Federica Pezzuto Journal: Front Oncol Date: 2022-08-04 Impact factor: 5.738