Objective:
To develop a new method for segmenting T1-Weighted magnetic resonance image into gray matter (GM), white matter (WM), cerebrospinal fluid (CSF), and other types of tissues.
Methods:
Our method has 4 steps. First, we divide each slice into overlapping blocks. Second, we fit thin plate splines to each block with different number of knots, and search for the knots configuration that gives us the smallest GCV score with a fudge factor [1]. Third, we fit the thin plate spline with the knots configuration found, and predict on a really fine grid. We also find the thresholds on the block with the K-means algorithm. Finally, we blend the predicted block images and the thresholds with some smooth weighting functions. The averages of the corresponding thresholds on all the blocks and the thresholds on the blended smooth image are calculated as well. We can apply all 3 thresholding scheme to the blended image, and pick the one that suits our needs best. Our empirical results show that images from different sources might need different thresholding.
The TPS method was tested on 5 subjects from the IBSR 20 normal data. We selected the 2nd, 6th, 10th, 14th, 18th subject after we sorted the 20 subjects based on their id′s. One slice near the middle of the brain for each subject was used with our method. SPM segmentation was applied to the whole volumes of the 5 subjects. Note that the 20 normal data is provided with human segmentation results.
Results & Discussion:
The comparison between TPS method, SPM, and manual segmentation on one subject is given in Figure 1 and 2. They show that TPS method matches manual segmentation reasonably well. And the TPS is doing really close to SPM. The correlation coefficients and kappa indices [2] between each 2 of the 3 methods for all the subjects were calculated (Table 1). The numbers confirm our conclusions. Note that the TPS method gives subpixel segmentation (Figure 3).
Conclusions:
Our method is an intensity based method. It uses thin plate splines to reconstruct the image. Our results show that the new method is comparable to the SPM segmentation, as well as the human segmentation. The TPS method has the advantage of giving subpixel level results and generating smoother boundaries. The partial volume effects are addressed by the subpixel segmentation. Also, it offers the potential for more accurate estimate of distances and curvatures on the cortical surfaces. Our method tackles the image non-uniformity through local thresholding and blending. It depends less heavily on correction for non-uniformity.
References & Acknowledgements:
[1] Luo, Z. and Wahba, G. (1997). Journal of the American Statistical Association, 92:107-116.
[2] Zijdenbos, A.P. et al (1994). IEEE Transactions on Medical Imaging, 13:716-724. 数据挖掘实验室
Figure 1: Comparison of TPS Segmentation vs Manual Segmentation 数据挖掘研究院
Figure 2: Pairwise Comparison of TPS, SPM, and Manual Segmentation
Figure 3: Zoomed View of the TPS Segmentation Boundaries
| Table 1: Coefficients for all the Comparisons with Mean and SD Summary | |||||
| subject no. |
corr. coef. | kappa index | |||
| GM | WM | GM | WM | ||
| tps vs manual | 1 | 0.660 | 0.827 | 0.836 | 0.872 |
| 2 | 0.702 | 0.757 | 0.841 | 0.827 | |
| 3 | 0.654 | 0.787 | 0.811 | 0.850 | |
| 4 | 0.410 | 0.678 | 0.723 | 0.770 | |
| 5 | 0.612 | 0.791 | 0.776 | 0.838 | |
| mean(sd) | |
0.608(0.115) | 0.768(0.056) | 0.798(0.049) | 0.831(0.038) |
| spm vs manual | 1 | 0.675 | 0.846 | 0.883 | 0.866 |
| 2 | 0.686 | 0.839 | 0.887 | 0.880 | |
| 3 | 0.637 | 0.810 | 0.863 | 0.842 | |
| 4 | 0.091 | 0.672 | 0.679 | 0.753 | |
| 5 | 0.450 | 0.803 | 0.825 | 0.824 | |
| mean(sd) | 数据挖掘实验室 |
0.518(0.250) | 0.794(0.071) | 0.827(0.087) | 0.833(0.050) |
| tps vs spm | 1 | 0.806 | 0.883 | 0.848 | 0.900 |
| 2 | 0.626 | 0.759 | 0.794 | 0.824 | |
| 3 | 0.734 | 0.822 | 0.808 | 0.861 | |
| 4 | 0.426 | 0.767 | 0.730 | 0.793 | |
| 5 | 0.645 | 0.800 | 0.785 | 0.836 | |
| mean(sd) | 数据挖掘研究院 |
0.647(0.143) | 0.806(0.050) | 0.793(0.043) | 0.843(0.040) |
