Radiomics and machine learning for characterizing CKD-associated cortical bone texture patterns in HR-pQCT tibia scans: a slice- and patient-level methodological framework.
Authors
Affiliations (6)
Affiliations (6)
- Weldon School of Biomedical Engineering, Purdue University, West Lafayette, IN 47907, United States.
- University Health Network, Toronto General Hospital, Toronto, ON M5G 2C4, Canada.
- School of Polymer Science and Engineering, Chonnam National University, Gwangju, 61186, Republic of Korea.
- Division of Nephrology and Hypertension, Indiana University School of Medicine, Indianapolis, IN 46202, United States.
- Department of Physical Therapy, School of Health and Human Sciences, Indiana University Indianapolis, Indianapolis, IN, United States.
- Departments of Endocrinology and Medicine, Austin Health, The University of Melbourne, Melbourne, VIC 3084, Australia.
Abstract
Chronic kidney disease (CKD) is associated with alterations in cortical bone structure and composition that contribute to increased fracture risk but are incompletely captured by standard clinical imaging, including DXA. This study evaluated whether radiomics combined with machine learning can enhance detection of CKD-related cortical bone characteristics using high-resolution peripheral quantitative computed tomography (HR-pQCT). HR-pQCT images (60.7 μm isotropic resolution; 168-slice stacks acquired at 7.3% and 30% proximal to the tibial endplate) were analyzed from 72 participants (38 non-CKD controls and 34 individuals with advanced CKD), yielding 24 192 cortical bone image slices. Cortical bone was segmented using a pretrained neural network optimized through transfer learning. Radiomic features were extracted using gray-level co-occurrence matrix (GLCM), local binary pattern (LBP), and combined GLCM+LBP feature sets, and paired with 7 machine-learning classifiers to systematically evaluate 21 feature-classifier combinations at distal and diaphyseal tibial sites. To address optimism associated with correlated slice-level data, radiomic features were aggregated per patient prior to model training and evaluated using patient-level splits (~7 test patients); this patient-level analysis constitutes the primary evaluation framework of this study. Under patient-level evaluation, classification performance was more variable across feature-classifier combinations and tibial sites, with wide CIs reflecting the limited sample size. At the slice level, hybrid GLCM+LBP features combined with XGBoost demonstrated the strongest classification performance across both tibial regions (distal AUC = 0.999; diaphyseal AUC = 0.999), providing methodological context for the systematic evaluation of all 21 feature-classifier combinations. Radiomics-derived texture features identified cortical heterogeneity that was not consistently reflected by conventional HR-pQCT metrics. These findings support radiomics as a promising methodological framework for characterizing CKD-associated cortical bone alterations while highlighting the importance of larger patient-level validation studies.