Xia Wu1, Peng Hu1, Liye Chen1, Yaoying Zhong1, Yudong Lin2, Xiaojing Yu1, Xi Hu1, Xinwei Tao3, Shushen Lin4, Tianye Niu5,6, Ran Chen7, Jihong Sun8,9. 1. Department of Radiology, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, 3 East Qingchun Road, Hangzhou, 310016, Zhejiang, China. 2. Zhejiang University School of Medicine, Hangzhou, 310011, China. 3. Bayer HealthCare, No.399, West Haiyang Road, Shanghai, China. 4. Siemens Healthineers China, No.399, West Haiyang Road, Shanghai, China. 5. Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, 310016, Zhejiang, China. 6. Institute of Translational Medicine, Zhejiang University, Hangzhou, 310016, Zhejiang, China. 7. Department of Diagnostic Ultrasound and Echocardiography, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, 310016, Zhejiang, China. 8. Department of Radiology, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, 3 East Qingchun Road, Hangzhou, 310016, Zhejiang, China. sunjihong@zju.edu.cn. 9. Cancer Center, Zhejiang University, Hangzhou, 310058, Zhejiang, China. sunjihong@zju.edu.cn.
Abstract
OBJECTIVES: To investigate the effects of slice thickness on CT radiomics features and models for staging liver fibrosis. METHODS: A total of 108 pathologically confirmed liver fibrosis patients from a single center were retrospectively collected and divided into different groups. Both thick (5- or 7-mm) and thin slices (1.3- or 2-mm) were analyzed. A fivefold cross-validation with 100 repeats was conducted. The minimum redundancy-maximum relevance algorithm was used to reduce the radiomics features, and the top 10 ranking features were included for further analysis for each loop. The random forest was used for model establishment. The models with median AUC were selected for the assessment of the discriminative performance for both datasets. Mutual features selected by the models with AUC > 0.8 were searched and considered as the most predictive ones. RESULTS: A total of 162 and 643 radiomics features with excellent reliability were selected from thick- and thin-slice datasets, respectively. The overall discriminative performance of the 500 AUCs from the thin-slice dataset was better than the thick slice. The median AUC values of the thick-sliced datasets were significantly lower than those of the thin-sliced datasets (0.78 and 0.90 for differentiating F1 vs. F2-4, 0.72 and 0.85 for differentiating F1-2 vs. F3-4, both P = 0.03). For differentiating F1-3 vs. F4, no significant difference was found (0.85 vs 0.94, P = 0.15). Six mutual predictive features across all the datasets were found. CONCLUSIONS: The radiomics features extracted from thin-slice images and their corresponding models were better and more stable for staging liver fibrosis.
OBJECTIVES: To investigate the effects of slice thickness on CT radiomics features and models for staging liver fibrosis. METHODS: A total of 108 pathologically confirmed liver fibrosis patients from a single center were retrospectively collected and divided into different groups. Both thick (5- or 7-mm) and thin slices (1.3- or 2-mm) were analyzed. A fivefold cross-validation with 100 repeats was conducted. The minimum redundancy-maximum relevance algorithm was used to reduce the radiomics features, and the top 10 ranking features were included for further analysis for each loop. The random forest was used for model establishment. The models with median AUC were selected for the assessment of the discriminative performance for both datasets. Mutual features selected by the models with AUC > 0.8 were searched and considered as the most predictive ones. RESULTS: A total of 162 and 643 radiomics features with excellent reliability were selected from thick- and thin-slice datasets, respectively. The overall discriminative performance of the 500 AUCs from the thin-slice dataset was better than the thick slice. The median AUC values of the thick-sliced datasets were significantly lower than those of the thin-sliced datasets (0.78 and 0.90 for differentiating F1 vs. F2-4, 0.72 and 0.85 for differentiating F1-2 vs. F3-4, both P = 0.03). For differentiating F1-3 vs. F4, no significant difference was found (0.85 vs 0.94, P = 0.15). Six mutual predictive features across all the datasets were found. CONCLUSIONS: The radiomics features extracted from thin-slice images and their corresponding models were better and more stable for staging liver fibrosis.