TriKey AutoRefine: A Stability-Aware Framework Built on Pan-organ Segmentation Radiomics for Chronic Kidney Disease Staging.
Authors
Affiliations (7)
Affiliations (7)
- School of Mechanical Science and Engineering, Huazhong University of Science and Technology, Wuhan, 430074, People's Republic of China.
- Department of Radiology, Tongji Hospital of Tongji Medical College of Huazhong University of Science and Technology, Wuhan, 430030, China.
- Department of Radiology, Renmin Hospital of Wuhan University, Wuhan, 430060, China.
- Department of Radiology, Hubei No.3, People's Hospital of Jianghan University , Wuhan, 430033, China.
- Department of Radiology, Tongji Hospital of Tongji Medical College of Huazhong University of Science and Technology, Wuhan, 430030, China. [email protected].
- School of Mechanical Science and Engineering, Huazhong University of Science and Technology, Wuhan, 430074, People's Republic of China. [email protected].
- Hubei Modern Manufacturing Quality Laboratory, School of Mechanical Engineering, Hubei University of Technology, Wuhan, 430068, China.
Abstract
Accurate and generalizable imaging‑based staging of chronic kidney disease remains challenging due to segmentation variability, feature instability, and heterogeneous clinical presentations. To address these gaps, a clinically aligned pan‑organ radiomics workflow for multi‑class chronic kidney disease staging was proposed. A total of 587 subjects from 3 centers were included, with chest and abdominal computed tomography scans paired with clinical and body-composition measurements. Segmentation‑induced variability was quantified within the pan‑organ segmentation system, revealing systematic anatomical and feature‑level patterns that enabled the construction of a robust and generalizable feature foundation. Building on these observations, the TriKey autorefine framework was introduced as a structured refinement strategy incorporating Shapley additive explanations guided refinement, clinically informed class weighting, and automatic parameter optimization. Using this framework, strong performance was achieved across chronic kidney disease staging tasks. In the standard 5‑class setting, high discriminative ability was obtained: training macro area under the curve 0.925 (95% confidence interval [CI] 0.923-0.927); validation macro area under the curve 0.875 ± 0.0087 (95% CI 0.858-0.892). In the clinically actionable 4‑class grouping, even stronger performance was observed, with a mean macro area under the curve of 0.915 ± 0.0060 (95% CI 0.903-0.927). Interpretability was further supported through radiomics-clinical correlation analyses, which confirmed that the selected pan organ features correspond to measurable physiological variations associated with chronic kidney disease progression. This study demonstrates a robust, interpretable, and clinically grounded pan‑organ radiomics pipeline for chronic kidney disease staging, providing a generalizable foundation for future pan-organ radiomics research in nephrology and broader clinical imaging applications.