Framework Results

Figure 8.4 presents one slice from an MRI scan of a mouse embryo, and an isosurface model of its liver extracted from the unprocessed dataset. Figure 8.5 presents 3D renderings of the sequence of steps performed on the mouse MRI data to segment the liver. The first step is the initialization, which includes smoothing the input data, thresholding followed by a a flood fill to remove isolated holes, and finally applying morphological operators to remove small gaps and protrusions on the surface. The second (surface deformation) step

Figure 8.4: (Left) one slice of a 256 x 128 x 128 MR scan of a mouse embryo. The central dark structure is its liver. (Right) a dual-threshold surface rendering highlights the segmentation problem.

first involves fitting to discrete edges and then to the gradient magnitude. This produces a significant improvement over the result in Fig. 8.4. Figure 8.8(a) presents several other structures that were segmented from the mouse embryo dataset. The skin (gray) and the liver (blue) were isolated using computational initialization. The brain ventricles (red) and the eyes (green) were segmented with interactive initialization.

The same set of initialization and surface deformation steps may be combined to extract a model of a spiny dendrite from the transmission electron microscopy (TEM) scan presented in Fig. 8.6(a). An isosurface extracted from the scan is presented in Fig. 8.6(b). Figure 8.7 shows the results of the proposed method compared to the results of a manual segmentation, which took approximately 10 hours of slice-by-slice hand contouring. The manual method suffers from slice-wise artifacts, and, because of the size and complexity of the dataset, the manual segmentation is unable to capture the level of detail that we obtain with the surface-fitting results. Manual segmentation can, however, form connections that are not well supported by the data in order to complete

Figure 8.5: (Left) the initialization of a mouse liver dataset using morphology to remove small pieces and holes. (Center) surface fitting to discrete edges. (Right) the final fit to maxima of gradient magnitude.

Figure 8.6: (a) One slice of a 154 x 586 x 270 TEM scan of a spiny dendrite shows low contrast and high noise content in a relatively complex dataset. (b) An isosurface rendering, with prefiltering, shows how noise and inhomogeneities in density interfere with visualizing the 3D structure of the dendrite.

Figure 8.6: (a) One slice of a 154 x 586 x 270 TEM scan of a spiny dendrite shows low contrast and high noise content in a relatively complex dataset. (b) An isosurface rendering, with prefiltering, shows how noise and inhomogeneities in density interfere with visualizing the 3D structure of the dendrite.

Figure 8.7: (Top) rendering of a dendrite segmented using our proposed method. (Bottom) rendering of a manual segmentation of the same dendrite.

the "spines" that cover this dendrite. These types of "judgments" that humans make when they perform such tasks by hand are a mixed blessing. Humans can use high-level knowledge about the problem to fill in where the data is weak, but the expectations of a trained operator can interfere with seeing unexpected or unusual features in the data.

Figure 8.8(c) presents models from four samples of an MR series of a developing frog embryo. The top left image (hour 9) shows the first evident structure, the blastocoel, in blue, surrounded by the outside casing of the embryo in gray.

(a)

Figure 8.8: (a) Final mouse embryo model with skin (gray), liver (blue), brain ventricles (red), and eyes (green). (b) Hour 16 dataset. (c) Geometric structures extracted from MRI scans of a developing frog embryo, with blastocoel (blue), blastoporal lip (red), and archenteron (green). Hour 9 (top left), hour 16 (top right), hour 20 (bottom left), and hour 30 (bottom right).

Table 8.1: Parameters for processing example datasets

Dataset

Initialization

Surface Fitting

Dendrite

1. Gaussian blur a = 0.5

1. Edge fitting: a = 0.75, threshold = 6, 3 =

= 0.1

2. Threshold: I < 127

2. Gradient magnitude fitting: a = 0.5, ¡3 =

1.0

3. Fill isolated holes

4. Morphology: 00.5 o C15

Mouse

1. Gaussian blur a = 0.5

1. Edge fitting: a = 0.75, threshold = 20, 3

= 2

2. Threshold: I > 3, I < 60

2. Gradient magnitude fitting: a = 0.5, ¡3 =

16.0

3. Fill isolated holes

4. Morphology: 02.0 o C3.0

Frog

1. Interactive

1. Gradient magnitude fitting: a = 1.25, 3 =

1.0

The top right image (hour 16) demonstrates the expansion of the blastocoel and the development of the blastoporal lip in red. In the bottom left image (hour 20) the blastoporal lip has collapsed, the blastocoel has contracted, and the archenteron in green has developed. In the bottom right image (hour 30) the blastocoel has collapsed and only the archenteron is present. For this dataset it was difficult to isolate structures only based on their voxel values. We therefore used our interactive techniques to isolate (during initialization) most of the structures in the frog embryo samples.

Table 8.1 describes for each dataset the specific techniques and parameters we used for the results in this section. These parameters were obtained by first making a sensible guess based on the contrasts and sizes of features in the data and then using trial and error to obtain acceptable results. Each dataset was processed between four and eight times to achieve these results. More tuning could improve things further, and once these parameters are set, they work moderately well for similar modalities with similar subjects. The method is iterative, but the update times are proportional to the surface area. On an SGI 180 MHz MIPS 10000 machine, the smaller mouse MR dataset required approximately 10 min of CPU time, and the dendrite dataset ran for approximately 45 min. Most of this time was spent in the initialization (which requires several complete passes through the data) and in the edge detection. The frog embryo datasets needed only a few minutes of processing time, because they did not require computational initialization and are significantly smaller than the other example datatsets.

Was this article helpful?

0 0

Post a comment