A deep learning system for automated kidney stone detection and volumetric segmentation on noncontrast CT scans.

Purpose: Early detection and size quantification of renal calculi are important for optimizing treatment and preventing severe kidney stone disease. Prior work has shown that volumetric measurements of kidney stones are more informative and reproducible than linear measurements. Deep learning-based systems that use abdominal noncontrast computed tomography (CT) scans may assist in detection and reduce workload by removing the need for manual stone volume measurement. Prior to this work, no such system had been developed for use on noisy low-dose CT or tested on a large-scale external dataset.

Methods: We used a dataset of 91 CT colonography (CTC) scans with manually marked kidney stones combined with 89 CTC scans without kidney stones. To compare with a prior work half the data was used for training and half for testing. A set of CTC scans from 6185 patients from a separate institution with patient-level labels were used as an external validation set. A 3D U-Net model was employed to segment the kidneys, followed by gradient-based anisotropic denoising, thresholding, and region growing. A 13 layer convolutional neural network classifier was then applied to distinguish kidney stones from false positive regions.

Results: The system achieved a sensitivity of 0.86 at 0.5 false positives per scan on a challenging test set of low-dose CT with many small stones, an improvement over an earlier work that obtained a sensitivity of 0.52. The stone volume measurements correlated well with manual measurements ( ). For patient-level classification, the system achieved an area under the receiver-operating characteristic of 0.95 on an external validation set (sensitivity = 0.88, specificity = 0.91 at the Youden point). A common cause of false positives were small atherosclerotic plaques in the renal sinus that simulated kidney stones.

Conclusions: Our deep-learning-based system showed improvements over a previously developed system that did not use deep learning, with even higher performance on an external validation set.