Renal corticomedullary segmentation from arterial phase volumetric T1-weighted imaging: human or machine?

Motivation: Renal corticomedullary volumes derived from arterial-phase fat-suppressed volumetric T1 weighted MRI have value in evaluation of kidney disease but manual segmentation is challenging. Goal(s): To assess performance of automatic versus manual segmentation of corticomedullary volumes. Approach: In 24 subjects, threshold-based and deep learning methods were used to automatically segment cortex and medullary volumes, which were compared to reference manual segmentation and a second reader. Results: Mean cortex and medullary Dice similarity coefficients of 0.91 and 0.95 for threshold segmentation (n=24), and 0.97 and 0.93 for deep learning segmentation (hold out test set, n=4) and 0.96 and 0.92 for Reader 2 (n=3) were obtained. Impact: MRI-derived renal corticomedullary segmentation can be efficiently and reproducibly performed using automated techniques with similar results to manual segmentation. Such techniques have promise for assessment and monitoring of chronic kidney disease and have potential application for prognostication through multi-parametric approaches.