Computerized FMD Image Analysis

Artery vasodilation assessment is a complex task owing to the poor quality of US image sequences and the small range of the vasodilation that has to be measured. Previous attempts to solve this problem were based on detecting the edges of the arterial wall. These methods have been successful to some extent; however, edge detection in ultrasound is prone to fail due to the presence of speckle

Table 5.9: Pearson's correlation coefficient, r (and p values), between serum lipid levels and the first five EigenD mode coefficients

Mode #

Cholesterol

Triglycerides

HDL-C

LDL-C

EigenD 1

—G.GG5 (G.949)

—G.217 (G.GG6)*

G.124 (G.117)

G.GS2 (G.684)

EigenD 2

—G.1G9 (G.169)

—G.G2S (G.774)

G.G96 (G.226)

—G.2S1 (G.GGS)*

EigenD S

—G.G72 (G.S66)

G.GG1 (G.992)

—G.128 (G.1G6)

—G.G12 (G.877)

EigenD 4

—G.G95 (G.228)

G.G58 (G.467)

G.G51 (G.517)

—G.145 (G.G68)

EigenD 5

G.G24 (G.76S)

G.147 (G.G62)

G.GG7 (G.9S4)

—G.G27 (G.7S5)

Table 5.10: Pearson's correlation coefficient, r (and p values), between traditional FMD indexes and the first five EigenFMD mode coefficients

Mode #

AFMDc

0basal

EigenFMD 1

-0.786 (<0.001)*

0.118 (0.136)

EigenFMD 2

-0.417 (<0.001)*

0.294 (<0.001)*

EigenFMD 3

-0.319 (<0.001)*

0.076 (0.339)

EigenFMD 4

-0.352 (<0.001)*

-0.190 (0.016)*

EigenFMD 5

0.113 (0.153)

-0.127 (0.108)

Statistically significant correlations are indicated with an asterisk.

Statistically significant correlations are indicated with an asterisk.

noise, poor quality edge definition and acoustic shadows. In our opinion, these techniques are based on low-level features with a poor reliability.

Our method, on the contrary, deals with the images in a more global manner. We model vasodilation as a scaling factor between frames that can be recovered by means of image registration techniques. The effect of low-level artifacts is therefore minimized as the registration measure is computed using all the information present in the whole image, and not just at the edges.

Results obtained with the automated method were better than those measured manually by medical experts. The proposed method presents a negligible bias (0.05 %FMD) whereas the bias of the manual measurements depends on the observer (range -0.16 to +0.34 %FMD). The standard deviation of the differences between the automated and the gold-standard measurements is 1.05 %FMD, which is slightly lower than the intra- and interobserver variabilities of manual measurements (1.13 %FMD and 1.20 %FMD, respectively). From the

Table 5.11: Pearson's correlation coefficient, r (and p values), between traditional FMD indexes and the first five EigenD mode coefficients

Table 5.11: Pearson's correlation coefficient, r (and p values), between traditional FMD indexes and the first five EigenD mode coefficients

Mode #

AFMDc

0basal

EigenD 1

0.170 (0.031)*

—0.987 (<0.001)*

EigenD 2

0.322 (<0.001)*

—0.053 (0.507)

EigenD 3

-0.479 (<0.001)*

0.062 (0.437)

EigenD 4

0.479 (<0.001)*

0.005 (0.952)

EigenD 5

0.263 (0.001)*

0.066 (0.409)

Statistically significant correlations are indicated with an asterisk.

Statistically significant correlations are indicated with an asterisk.

dilation CV, the proposed method has also shown to present better reproducibility (CV = 0.40%) than the manual procedure (CV = 1.04%).

The method is reasonably fast. Our experiments were carried out on a PC (Pentium III @ 600 MHz) under RedHat 7.2 Linux operating system. The registration algorithm and the Kalman filtering are both coded in C++ without a thorough code optimization since the implementation of the registration method is a general-purpose software not specifically devised for this application. Under these constraints, the mean execution times per frame are 6.4 sec (SD = 0.8 sec) and 4.0 sec (SD = 1.2 sec) for the first and second phase, respectively. This time also incorporates outputting of progress information. From our experience with the software, we think that these figures could still be cut down substantially by customizing and further optimizing several parts of the code.

The vasodilation model used in this approach has also some potential limitations. Here, dilation is recovered by means of reduced similarity transformation between each frame and the reference one. However, this implicitly assumes that the wall thickness dilates in the same way that the artery does, while it may remain constant or even thin during lumen dilation. The unstable presence of the lumen-intima boundary (LIB) could potentially affect the registration results. Finally, structures stuck to the outer part of the arterial wall may introduce errors in the vasodilation measurements since they make it more difficult to adequately pad the reference frame. The results obtained in this chapter seem to indicate, however, that the vasodilation model outlined in this work is a reasonable simplification.

Motion compensation is necessary to avoid potential sources of bias in the subsequent estimation of vasodilation and to ensure that vasodilation is measured by comparing the same artery segment in two different frames. Nevertheless, stable motion references are required to succeed in motion recovery and avoiding indetermination of the correct alignment in the longitudinal direction of the artery. Moreover, only 2-D information is available in the image to correct a problem that is intrinsically 3-D in nature.

Another advantage of motion compensation is that it makes unnecessary the manual [12,13] or automatic [15] tracking of a region of interest (ROI) in the image sequence. This ROI tracking is required for the edge detection of some of the methods proposed in the literature. Our technique requires a simple preprocessing of only the reference frame. The interaction required is minimal (only rough delineation of two lines) and introduces a small variability (it is included in the CV of 0.40% obtained in the reproducibility study).

The initialization of the registration algorithm is a very important aspect. This initial transformation should fall inside the capture range of the algorithm [22] whose size depends on many factors, and its determination is not possible a priori. Some of these factors are the image quality, the line thickness of the arterial walls and the preprocessing made to the images. It is common that some frames appear with poor image quality along the sequence due to patient's motion. One of these frames may probably lead to erroneous registration values. To reduce error propagation Kalman filtering has shown to be very valuable.

Finally, it is important to recall that the proposed tracking strategy depends on a model of mean arterial vasodilation. This model was estimated from a number of training vasodilation curves, which corresponded to young healthy volunteers. Therefore, this model could bias the analysis of sequences coming from a general population or in specific subject groups like old obese patients. This limitation could be overcome by using a larger training set for building the mean vasodilation model or by having several models for different age groups.

Was this article helpful?

0 0
Lower Your Cholesterol In Just 33 Days

Lower Your Cholesterol In Just 33 Days

Discover secrets, myths, truths, lies and strategies for dealing effectively with cholesterol, now and forever! Uncover techniques, remedies and alternative for lowering your cholesterol quickly and significantly in just ONE MONTH! Find insights into the screenings, meanings and numbers involved in lowering cholesterol and the implications, consideration it has for your lifestyle and future!

Get My Free Ebook


Post a comment