Article Figures & Data
Figures
- fig 1.
Examples of MR images that may be used for segmentation.
A, Contrast-enhanced T1-weighted image (500/8/2).
B, T2-weighted image (6000/14/2).
- fig 2.
This T1-weighted image (500/8/2) illustrates the selection of tag points by an operator. The accuracy of segmentation is dependent on tag points representing the various tissue types being segmented
- fig 3.
The segmentation maps illustrated in this figure are representative of the types of maps generated by the nearest-neighbor algorithm. They were taken at a variety of levels from more than one patient.
A, White matter.
B, Gray matter.
C, CSF.
D, Tumor contrast enhancement (arrow).
E, Edema (arrow).
F, Cystic compartment (arrow).
- fig 4.
A, This contrast-enhanced T1-weighted image (500/8/2) illustrates the tumor boundary defined by an operator as part of the surface modeling algorithm approach to volume analysis.
B, This figure illustrates the process of generating the manually defined standard of reference. An operator overlays a chosen color on a specific tissue type (contrast-enhancing tumor tissue in the case of this study). Accuracy is improved by varying the image and color overlay intensity and magnitude. Despite techniques to improve accuracy, the process remains subjective.
- fig 5.
The basic principles behind the surface modeling algorithm, which uses a parametric mesh approach, are as follows: an operator defines the structure of interest by tracing points on a 2D MR slice, the algorithm uniformly redigitizes the points at each level of the image for the region of interest, the points from adjacent sections are then sewn together using triangular tiles, and the result is a tiled parametric mesh model from which volumes are calculated
- fig 6.
This figure shows the change in contrast enhancement and the corresponding growth rates computed from a manually defined standard over a 250-day period in a 44-year-old subject. By generating volumetric data across time, the patient may be tracked and response or nonresponse to therapy documented
- fig 7.
A, The nearest-neighbor generated enhancement volumes (cm3) are highly correlated (r2 = .99) with the manually defined standard. Though not able to duplicate the volumes as determined manually, the nearest-neighbor algorithm is systematic and accurate in its ability to quantify tumor volume based on contrast-enhancing tissue volumes.
B, The surface modeling algorithm generated enhancement volumes (cm3) are also highly correlated (r2 = .94) with the manually defined standard. It is unable to separate nonenhancing necrotic areas from surrounding enhancing areas and cannot generate volumes based on noncontiguous lesions. Therefore, the surface modeling algorithm has a tendency to overestimate and underestimate enhancement volumes to a greater degree than does the nearest-neighbor algorithm.
- fig 8.
A, Growth rates (in days) measured in terms of halving times and doubling times for the nearest-neighbor algorithm are highly correlated with the growth rates for the manually defined volumes (r2 = .96).
B, Growth rates (in days) generated from the surface modeling algorithm were not highly correlated with growth rates generated from the manually defined standard of reference (r2 = .45).
Tables
Comparison of nearest-neighbor segmentation, surface modeling, and the manually defined standard of reference in terms of operator input and accuracy
