Deep learning body-composition analysis of clinically acquired CT-scans estimates creatinine excretion with high accuracy in patients and healthy individuals.

Abstract

Assessment of daily creatinine production and excretion plays a crucial role in the estimation of renal function. Creatinine excretion is estimated by creatinine excretion equations and implicitly in eGFR equations like MDRD and CKD-EPI. These equations are however unreliable in patients with aberrant body composition. In this study we developed and validated equations estimating creatinine production using deep learning body-composition analysis of clinically acquired CT-scans. We retrospectively included patients in our center that received any CT-scan including the abdomen and had a 24-h urine collection within 2 weeks of the scan (n = 636). To validate the equations in healthy individuals, we included a kidney donor dataset (n = 287). We used a deep learning algorithm to segment muscle and fat at the 3rd lumbar vertebra, calculate surface areas and extract radiomics parameters. Two equations for CT-based estimate of RenAl FuncTion (CRAFT 1 including CT parameters, age, weight, and stature and CRAFT 2 excluding weight and stature) were developed and compared to the Cockcroft-Gault and the Ix equations. CRAFT1 and CRAFT 2 were both unbiased (MPE = 0.18 and 0.16 mmol/day, respectively) and accurate (RMSE = 2.68 and 2.78 mmol/day, respectively) in the patient dataset and were more accurate than the Ix (RMSE = 3.46 mmol/day) and Cockcroft-Gault equation (RMSE = 3.52 mmol/day). In healthy kidney donors, CRAFT 1 and CRAFT 2 remained unbiased (MPE = - 0.71 and - 0.73 mmol/day respectively) and accurate (RMSE = 1.86 and 1.97 mmol/day, respectively). Deep learning-based extraction of body-composition parameters from abdominal CT-scans can be used to reliably estimate creatinine production in both patients as well as healthy individuals. The presented algorithm can improve the estimation of renal function in patients who have recently had a CT scan. The proposed methods provide an improved estimation of renal function that is fully automatic and can be readily implemented in routine clinical practice.