DTI-based radiomics signature for the detection of early diabetic kidney damage

Yi Deng,Bi-ran Yang,Jin-wen Luo,Guo-xin Du,Liang-ping Luo
DOI: https://doi.org/10.1007/s00261-020-02576-6
IF: 2.4
2020-01-01
Abdominal Radiology
Abstract:Objective To explore whether a radiomics signature based on diffusion tensor imaging (DTI) can detect early kidney damage in diabetic patients. Materials and methods Twenty-eight healthy volunteers (group A) and thirty type 2 diabetic patients (group B) with micro-normoalbuminuria, a urinary albumin-to-creatinine ratio (ACR) < 30 mg/g and an estimated glomerular filtration rate (eGFR) of 60–120 mL/(min 1.73 m 2 ) were recruited. Kidney DTI was performed using 1.5T magnetic resonance imaging (MRI).The radiologist manually drew regions of interest (ROI) on the fractional anisotropy (FA) map of the right kidney ROI including the cortex and medulla. The texture features of the ROIs were extracted using MaZda software. The Fisher coefficient, mutual information (MI), and probability of classification error and average correlation coefficient (POE + ACC) methods were used to select the texture features. The most valuable texture features were further selected by the least absolute shrinkage and selection operator (LASSO) algorithm. A LASSO regression model based on the radiomics signature was established. The diagnostic performance of the model for detecting early diabetic kidney changes was evaluated by the area under the receiver operating characteristic (ROC) curve (AUC). Empower (R), R, and MedCalc15.8 software were used for statistical analysis Results A total of 279 texture features were extracted from ROI of the kidney, and 30 most valuable texture features were selected from groups A and B using MaZda software. After LASSO-logistic regression, a diagnostic model of diabetic kidney damage based on texture features was established. Model discrimination evaluation: AUC = 0.882 (0.770 ± 0.952). Model calibration evaluation: Hosmer–Lemeshow X2 = 5.3611, P = 0.7184, P > 0.05, the model has good calibration. Conclusion The texture features based on DTI could play a promising role in detecting early diabetic kidney damage.
What problem does this paper attempt to address?