Quantitative evaluation of chronically obstructed kidneys from noncontrast computed tomography based on deep learning

To quantitatively report renal parenchymal volume (RPV), renal sinus volume (RSV), and renal parenchymal density (RPD) for chronically obstructed kidneys from noncontrast computed tomography (NCCT).

This retrospective study was approved by the institutional review board of our hospital with a waiver of informed consent. We retrospectively collected 304 consecutive NCCT scans of urinary obstruction and constructed two datasets: one with 167 patient scans for parenchyma and sinus segmentation (segmentation dataset) and the other containing 137 scans from different patients diagnosed with chronic urinary obstruction (CUO dataset) and paired with split glomerular filtration rate (sGFR). A cascaded three-dimensional (3D) U-Net model was developed and validated for parenchyma and sinus segmentation. The RPV, RSV, and RPD of the CUO dataset were calculated by the model with manual editing. A multivariate analysis was performed to show the association between all parameters and the sGFR.

In the test dataset, the Dice values for parenchyma and sinus segmentation were 0.95 ± 0.04 and 0.90 ± 0.05, respectively. Compared with those of nonobstructed kidneys, the RSV and RPD of obstructed kidneys increased, but RPV and sGFR decreased (P < .001). For chronically obstructed kidneys, age (r = -0.292, P < .001), RPV (r = 0.849, P < .001), RSV (r = -0.331, P < .001), and RPD (r = -0.296, P < .001) were significantly correlated with sGFR. The fitted regression model was sGFR = 10.873-0.111 Age + 0.211 RPV - 0.022 RSV (r² = 0.712).

NCCT combined with deep learning has the potential to be a single radiological procedure for morphological and functional evaluation of chronically obstructed kidneys.