Automatic segmentation of polycystic kidneys from magnetic resonance images using decision tree classification and snake algorithm

Abstract

Autosomal dominant polycystic kidney disease (ADPKD) is characterized by progressive bilateral renal cyst formation, leading to severe increases in kidney volume and loss of function. Total kidney volume (TKV) is the only established biomarker for tracking ADPKD. This is measured multiple times per year from each patient to examine the extent of renal enlargement and overall cyst load. Currently this is conducted by planimetry tracing, which involves manually delineating kidneys from surrounding tissues in the abdominal cavity using a digital drawing tool. By performing this on every image in a magnetic resonance scan, TKV is estimated. This is a time-consuming and laborious process for radiologists. Our aim is to develop an automated method for ADPKD patient kidney segmentation and quantifying TKV. Thirteen MRI scans of kidneys ranging across the spectrum from normal to severe cyst load were analyzed. Images were separated into two halves, each made up of 200 square regions. Features were extracted from grayscale values of each region, and these data were combined in a supervised decision tree algorithm to classify between kidney and non-kidney regions. Filtering and dilation were applied to the classified 400x400 matrix in order to roughly segment the kidneys. Contrast enhancement and k-means clustering was performed before applying an active contour function to determine kidney edges. Eccentricity analysis confirmed appropriate relative sphericity for segmented kidney shapes, before combining their areas with linear extrapolation to estimate TKV. This protocol is evaluated against clinical reference standard TKV measurements.