Obstructive nephropathy (ON) is a disease in which the renal parenchyma is pathologically and functionally damaged due to changes in the structure or function of the urinary system, resulting in obstruction of urine excretion.[1] The main causes of urinary tract obstruction are the ureter itself (such as calculi and stricture deformity) and outside the ureter (cyst and tumor mass). [2] The clinical diagnosis of ON requires the help of some auxiliary examinations. The standard four-phase computed tomography urography (CTU) includes the unenhanced phase (UP), corticomedullary phase (CMP), nephrographic phase, and excretory phase, which can help clinicians identify the location and degree of obstruction of ON.[3] As the gold standard for precise assessment of renal function, emission computed tomography (ECT) can accurately evaluate renal function and the degree of damage, which is critical for the development of therapy and prognosis of ON.[4] Unfortunately, ECT is an intrusive technique that necessitates intravenous radiopharmaceutical imaging, which, in addition to lengthening the appointment time and raising the financial burden, can increase the stress on the kidneys, particularly in patients with poor renal function.[5]
The cortical part of the renal parenchyma is composed of more than one million nephrons and is the main place where the kidneys perform filtration. When obstructive hydrops or some space-occupying diseases occur, the morphology of the renal cortex will be compressed and thinned, resulting in a progressive decline in renal function. Clinically, experienced physicians can roughly judge the renal function of the affected side of ON by the specific morphology of the renal cortex on the CTU image. However, this strategy places more demands on physicians' professional knowledge and experience, and there are inter-observer variances and poor reproducibility.
Glomerular filtration rate (GFR) is an essential indicator representing renal filtration function, but accurate measurement of GFR remains a challenge in clinical practice.[6] Modification of Diet in Renal Disease (MDRD)[7, 8] and Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI)[9] equations are commonly used methods to estimate GFR in patients with chronic kidney disease. However, both methods have limitations in their application and have been reported to be misestimated in some circumstances, failing to cover all patient characteristics.[10, 11]
In recent years, deep convolutional neural networks (CNN) have shown excellent performance in many computer vision and biomedical applications, such as disease diagnosis [12], organ segmentation [13], and object detection [14]. Three-dimensional (3D) segmentation based on deep learning (DL) has been proven to be an extremely successful approach for kidney image segmentation.[15, 16] However, most studies are limited to a small amount of data and only focus on the performance improvement of the model, lacking practical clinical application scenarios.
In this study, we developed an end-to-end automatic prediction system for single-kidney function, named UroAngel, based on CNN of 3D U-Net, and achieved non-invasive, convenient, and reliable results. We used 400 CTU images from 100 patients for training and validation of the 3D U-Net model, and 1680 CTU images from 420 patients for constructing the logistic regression model. Finally, the accuracy of UroAngel with modified MDRD, CKD-EPI, and expert radiologists in predicting GFR was compared in 40 ON patients to validate the clinical effectiveness.