Improving the segmentation of therapy-induced leukoencephalopathy in children with acute lymphoblastic leukemia using a priori information and a gradient magnitude threshold

Abstract

Reliably quantifying therapy-induced leukoencephalopathy is a challenging task due to the similarity between its MR properties and those of normal tissues. Multispectral MR images were analyzed for 15 children treated for acute lymphoblastic leukemia. Three different analysis techniques were compared to examine improvements in the segmentation accuracy of leukoencephalopathy versus manual tracings by two experienced observers. The original technique used a white matter mask based on the segmentation of the first serial examination of each patient and no a priori information. The modified techniques combine spatially normalized a priori maps as input and a gradient magnitude threshold. The second technique used a 2D threshold, while the third algorithm utilized a 3D threshold. MR images were segmented with a Kohonen self-organizing map for all three algorithms. Kappa values were compared for the three techniques to each observer and statistically significant improvements were seen between the original and third algorithms (Observer 1: 0.651, 0.744, P = 0.015; Observer 2: 0.603, 0.699, P = 0.024). More accurate and reliable quantification reduces the amount of variance in MR measures and facilitates clinical trials to determine the clinical significance of leukoencephalopathy in this vulnerable population.

(c) 2004 Wiley-Liss, Inc.

Figures

FIG. 1

Example of the input images used in all three segmentation algorithms. The MR images are shown on the top row and from left to right correspond to the T1, T2, PD, and FLAIR-weighted images acquired for all subjects on the institutional protocol. The a priori maps spatially normalized to the patient’s MR examination are shown on the bottom row. From left to right, the white matter, gray matter, and cerebrospinal fluid a priori maps are shown that were used to generate the white matter mask and served as additional input to the second and third leukoencephalopathy segmentation algorithms.

FIG. 2

Schematic of the 3D kernel used in the third leukoencephalopathy algorithm. The pixel shown in gray is the reference pixel for which the gradient magnitude value is determined, while the pixels represented as black identify the neighbors used in the calculation of the gradient magnitude value. The center group of pixels represents the in-plane component of the kernel, while the left and right represent pixels in the MR slices superior and inferior to the reference pixel, respectively.

FIG. 3

Example output from the three segmentation algorithms. The first two images on the left are the manual segmentations overlaid on FLAIR image corresponding to the MR slice that was segmented. The first image on the left is from Observer 1 and the second is from Observer 2. The remaining three images correspond to the three segmentation algorithms. The center image is the output from the original leukoencephalopathy segmentation algorithm without a priori information or a gradient magnitude threshold. The second from the right is the second leukoencephalopathy segmentation output showing the improvements utilizing the a priori maps and a 2D gradient magnitude threshold. On the right, the improvements provided by the 3D gradient magnitude threshold are shown in the output from the third segmentation algorithm.

FIG. 4

Scatterplots of the volume of leukoencephalopathy from the manual and automated segmentation algorithms. The horizontal axes represent the volume of the manual segmentation for the single slice used in each of the 15 subjects reported in cubic centimeters. The graph on the top shows volumes from the manual segmentation of Observer 1 while the results from Observer 2 are shown on the bottom. Both vertical axes represent the corresponding single-slice volume from the original and third segmentation algorithms in cubic centimeters. The original technique volumes are shown by empty squares and the third technique is represented by filled squares. The dotted line represents the line of equality between the manual and automated techniques.