Performance of an automated renal segmentation algorithm based on morphological erosion and connectivity

摘要

The precision, accuracy, and efficiency of a novel semi-automated renal segmentation technique for volumetric interpolated breath-hold sequence (VIBE) MRI sequences was analyzed using 7 clinical datasets (14 kidneys). Two observers performed whole-kidney segmentation using Edge Wave segmentation software based on constrained morphological growth. Ground truths were prepared by manual tracing of kidney on each of approximately 40 slices. Using the software, the average inter-observer disagreement was 2.7%± 2.1% for whole kidney volume, 2.1%± 1.8% for cortex, and 4.1%± 3.2% for medulla. In comparison to the ground truth kidney volume, the error was 2.8%± 2.5% for whole kidney volume, 3.1%± 1.7% for cortex, and 3.6%±.3.1% for medulla. It took an average of 4:14±1:42 minutes of operator time, plus 2:56± 1:23 minutes of computer time to perform segmentation of one whole kidney, cortex, and medulla. Segmentation speed, inter-observer agreement and accuracy were several times better than those of our existing graph-cuts segmentation technique requiring approximately 20 minutes per case and with 7-10% error. With the expedited image processing, high inter-observer agreement, and volumes closely matching true values, kidney volumetry becomes a reality in many clinical applications.