A contrast CT scan is conducted on the donor candidate during kidney transplantation to detect subclinical abnormalities in the kidney. A manual image-processing method was used to assess kidney, cortex, and medulla volumes in the Aging Kidney Anatomy research. However, because the procedure is time intensive and inconvenient for clinical treatment, these measures are not acquired during donor assessments. For a study, researchers provided a completely automated segmentation method for assessing the volumes of the kidney, brain, and medulla.
The method was developed using 1,930 contrast-enhanced CT examinations with reference standard manual segmentations from a single institution. A convolutional neural network model was trained (n=1238), verified (n=306), and then tested against a collection of reference standard segmentations in a hold-out test set (n=386). Following the first evaluation, the method was evaluated on datasets from two external sites (n=1226).
The automated model performed similarly to manual segmentation, with errors comparable to interobserver variability in human segmentation. In the test set, the automated technique achieved Dice similarity metrics of 0.94 (right cortex), 0.90 (right medulla), 0.94 (left cortex), and 0.90 (left medulla) when compared to the reference standard. When the technique was applied to the two external datasets, similar results were obtained. A fully automated method for assessing cortical and medullary volumes in kidney CT images has been developed. The approach might be beneficial in a variety of therapeutic settings.