Bias Estimation Protocol

From section 9.5.2, we compute the error per vertex (point) around the boundary consisting of 150 points for MRF and FCM (with smoother) methods. We used the PDM ruler for bias error analysis. It can be seen from Fig. 9.39 that there is no bias error for the FCM method, and all the boundary points have an error with a mean of 1.5 pixels. For the MRF method, the bias error curve first becomes negative, and then rises to positive values after the boundary point 45. This shows that there has been a right shift of the computer-estimated contour compared to the ideal contour after the point 45. This also means that the one third of the contour is inside the ideal boundary, and the remaining two thirds of the contour is outside the ideal boundary. Such a behavior can be explained using two concepts: (a) intra- and interobserver variability and (b) shift in the estimated contour. This is out of the scope of this chapter and will be discussed elsewhere.

Figure 9.39: The bias error is compared between the MRF and the FCM with smoother. The mean errors are plotted against consecutive points around the contour. Large noise protocol.

e (pixels*2) Variance (plxetsA2)

Figure 9.40: PDM vs. SDM methods. Left: MRF: PDM vs. SDM. Right: FCM: PDMvs. SDM. The length of the range of the mean errors is less than 0.36 pixels, and the difference between the two curves is about 0.03 pixels.

e (pixels*2) Variance (plxetsA2)

Figure 9.40: PDM vs. SDM methods. Left: MRF: PDM vs. SDM. Right: FCM: PDMvs. SDM. The length of the range of the mean errors is less than 0.36 pixels, and the difference between the two curves is about 0.03 pixels.

0 0

Post a comment