Jun Shu1, Didi Wen1, Yibin Xi1, Yuwei Xia2, Zhengting Cai2, Wanni Xu3, Xiaoli Meng4, Bao Liu1, Hong Yin5. 1. Department of Radiology, Xijing Hospital, Fourth Military Medical University, Changle West Road 127#, Xi'an City, 710032, People's Republic of China. 2. Huiying Medical Technology Co., Ltd. Room C103, B2, Dongsheng Science and Technology Park, HaiDian District, Beijing City, 100192, People's Republic of China. 3. Department of Pathology, Xijing Hospital, Fourth Military Medical University, Changle West Road 127#, Xi'an City, 710032, People's Republic of China; Deng Road 97#, Xi'an City, 710077, People's Republic of China. 4. Department of Radiology, Xi'an XD Group Hospital, Shaanxi University of Chinese Medicine, Feng Deng Road 97#, Xi'an City, 710077, People's Republic of China. 5. Department of Radiology, Xijing Hospital, Fourth Military Medical University, Changle West Road 127#, Xi'an City, 710032, People's Republic of China. Electronic address: yinhong@fmmu.deu.cn.
Abstract
PURPOSE: To evaluate the performance of machine learning (ML)-based computed tomography (CT) radiomics analysis for discriminating between low grade (WHO/ISUP I-II) and high grade (WHO/ISUP III-IV) clear cell renal cell carcinomas (ccRCCs). METHODS: A total of 164 low grade and 107 high grade ccRCCs were retrospectively analyzed in this study. Radiomic features were extracted from corticomedullary phase (CMP) and nephrographic phase (NP) CT images. Intraclass correlation coefficient (ICC) was calculated to quantify the feature's reproducibility. The training and validation cohort consisted of 163 and 108 cases. Least absolute shrinkage and selection operator (LASSO) regression method was used for feature selection. The machine learning (ML) classifiers were k-NearestNeighbor (KNN), Logistic Regression (LR), multilayer perceptron (MLP), Random Forest (RF), and support vector machine (SVM). The performance of classifiers was mainly evaluated and compared by certain metrics. RESULTS: Seven CMP features (ICC range, 0.990-0.999) and seven NP features (ICC range, 0.931-0.999) were selected. The accuracy of CMP, NP and the combination of CMP and NP ranged from 82.2%-85.9 %, 82.8%-94.5 % and 86.5%-90.8 % in the training cohort, and 90.7%-95.4%, 77.8%-79.6 % and 91.7%-93.5 % in the validation cohort. The AUC of CMP, NP and the combination of CMP and NP ranged from 0.901 to 0.938, 0.912 to 0.976, 0.948 to 0.968 in the training cohort, and 0.957 to 0.974, 0.856 to 0.875, 0.960 to 0.978 in the validation cohort. CONCLUSIONS: ML-based CT radiomics analysis can be used to predict the WHO/ISUP grade of ccRCCs preoperatively.
PURPOSE: To evaluate the performance of machine learning (ML)-based computed tomography (CT) radiomics analysis for discriminating between low grade (WHO/ISUP I-II) and high grade (WHO/ISUP III-IV) clear cell renal cell carcinomas (ccRCCs). METHODS: A total of 164 low grade and 107 high grade ccRCCs were retrospectively analyzed in this study. Radiomic features were extracted from corticomedullary phase (CMP) and nephrographic phase (NP) CT images. Intraclass correlation coefficient (ICC) was calculated to quantify the feature's reproducibility. The training and validation cohort consisted of 163 and 108 cases. Least absolute shrinkage and selection operator (LASSO) regression method was used for feature selection. The machine learning (ML) classifiers were k-NearestNeighbor (KNN), Logistic Regression (LR), multilayer perceptron (MLP), Random Forest (RF), and support vector machine (SVM). The performance of classifiers was mainly evaluated and compared by certain metrics. RESULTS: Seven CMP features (ICC range, 0.990-0.999) and seven NP features (ICC range, 0.931-0.999) were selected. The accuracy of CMP, NP and the combination of CMP and NP ranged from 82.2%-85.9 %, 82.8%-94.5 % and 86.5%-90.8 % in the training cohort, and 90.7%-95.4%, 77.8%-79.6 % and 91.7%-93.5 % in the validation cohort. The AUC of CMP, NP and the combination of CMP and NP ranged from 0.901 to 0.938, 0.912 to 0.976, 0.948 to 0.968 in the training cohort, and 0.957 to 0.974, 0.856 to 0.875, 0.960 to 0.978 in the validation cohort. CONCLUSIONS: ML-based CT radiomics analysis can be used to predict the WHO/ISUP grade of ccRCCs preoperatively.
Authors: Claudia-Gabriela Moldovanu; Bianca Boca; Andrei Lebovici; Attila Tamas-Szora; Diana Sorina Feier; Nicolae Crisan; Iulia Andras; Mircea Marian Buruian Journal: J Pers Med Date: 2020-12-23