Chun-Qiang Lu1, Yuan-Cheng Wang1, Xiang-Pan Meng1, Hai-Tong Zhao2, Chu-Hui Zeng1, Weiwei Xu1, Ya-Ting Gao1, Shenghong Ju3. 1. Jiangsu Key Laboratory of Molecular and Functional Imaging, Department of Radiology, Zhongda Hospital, Medical School of Southeast University, 87 Ding Jia Qiao Road, Nanjing, 210009, China. 2. Lab of Image and Signal Processing, School of Biological Science and Medical Engineering, Southeast University, Nanjing, China. 3. Jiangsu Key Laboratory of Molecular and Functional Imaging, Department of Radiology, Zhongda Hospital, Medical School of Southeast University, 87 Ding Jia Qiao Road, Nanjing, 210009, China. jsh0836@hotmail.com.
Abstract
OBJECTIVES: To identify CT markers for screening of early type 2 diabetes and assessment of the risk of incident diabetes using a radiomics method. METHODS: The medical records of 26,947 inpatients were reviewed. A total of 690 patients were selected and allocated to a primary cohort, a validation cohort, and a prediction cohort and used to build prediction models for diabetes. Three radiomics signatures were constructed using CT image features extracted from three regions of interest, i.e., in the pancreas, liver, and psoas major muscle. By incorporating radiomics signatures and other markers, we built a radiomics nomogram that could be used to screen for early diabetes and predict future diabetes. RESULTS: Of the three abdominal organs for which radiomics signature were constructed, that of the pancreas showed the best discriminatory power for early diabetes screening and prediction (C-statistics of 0.833, 0.846, and 0.899 for the primary cohort, validation cohort, and prediction cohort, respectively). The sensitivity and specificity of the nomogram for prediction of 3-year incident diabetes were 0.827 and 0.807, respectively. CONCLUSIONS: This study presents alternative radiomics markers that have potential for use in screening for undiagnosed type 2 diabetes and prediction of 3-year incident diabetes. KEY POINTS: • CT images may provide useful information to evaluate the risk of developing diabetes. • Radiomics score for diabetes prediction is based on subtle changes of abdominal organs detected by CT. • The radiomics signature of pancreas, a combination of five features of CT images, is efficient for early diabetes screening and prediction of future diabetes (AUC > 0.8).
OBJECTIVES: To identify CT markers for screening of early type 2 diabetes and assessment of the risk of incident diabetes using a radiomics method. METHODS: The medical records of 26,947 inpatients were reviewed. A total of 690 patients were selected and allocated to a primary cohort, a validation cohort, and a prediction cohort and used to build prediction models for diabetes. Three radiomics signatures were constructed using CT image features extracted from three regions of interest, i.e., in the pancreas, liver, and psoas major muscle. By incorporating radiomics signatures and other markers, we built a radiomics nomogram that could be used to screen for early diabetes and predict future diabetes. RESULTS: Of the three abdominal organs for which radiomics signature were constructed, that of the pancreas showed the best discriminatory power for early diabetes screening and prediction (C-statistics of 0.833, 0.846, and 0.899 for the primary cohort, validation cohort, and prediction cohort, respectively). The sensitivity and specificity of the nomogram for prediction of 3-year incident diabetes were 0.827 and 0.807, respectively. CONCLUSIONS: This study presents alternative radiomics markers that have potential for use in screening for undiagnosed type 2 diabetes and prediction of 3-year incident diabetes. KEY POINTS: • CT images may provide useful information to evaluate the risk of developing diabetes. • Radiomics score for diabetes prediction is based on subtle changes of abdominal organs detected by CT. • The radiomics signature of pancreas, a combination of five features of CT images, is efficient for early diabetes screening and prediction of future diabetes (AUC > 0.8).