CAPÍTULO II...................................................................................................................................... 12
2. MARCO DE REFERENCIA
2.7. Identificación y operacionalización de las variables
Chapter 2, Section 2.4.1 mentioned the concept of achieving segmentation by analysis of the motion in the image. A technique proposed by Kiriyanthan et al. incorporated a motion segmentation within an iterative registration algorithm to preserve the motion discontinuity [102]. The group showed some preliminary success in abdominal images of the liver. While Kiriyanthan et al.’s work is directly relevant, it is still under development, can only cope with small displacements and would be difficult to reproduce. This was not the focus of this PhD and this route was not considered (particularly as the paper was published towards the end of this PhD). Instead, readily available tools were used to gain insight into the potential of using motion segmentation as a component of the visceral slide processing technique.
A picture of magnitude of motion can be obtained by registering each frame to its consecutive frame without imposing a mask. Both ANTs and ShIRT registration algorithms were used to
169 produce images of motion magnitude on the same cine-MRI slice. The ShIRT registration used the default parameters used throughout this thesis (node spacing of 4 and an adaptive smoothness constraint strength). The ANTs registration parameters used were:
- Cross correlation similarity metric with a regular sampling grid - Diffeomorphic transformation
- Multiscale approach with five levels
- No pre-rigid registration step (as the images are consecutive and already aligned)
Upon summation of the magnitude of motion between each frame, images were produced from the results from each of the registration algorithms, as shown in Figure 6.3.
(a) (b)
Figure 6.3: Images showing the summation of the displacement fields from registration of consecutive frames of a cine-MRI with: (a) diffeomorphic registration performed using ANTs and (b) default parameters in ShIRT
ANTs was selected as the primary tool in this investigation as relaxation of the smoothness constraint was expected to permit a sharper change in displacement and therefore a better defined cut-off at the edge of the moving region. This can be seen to be the case when comparing the ANTs registration (Figure 6.3a) to the registration using ShIRT (Figure 6.3b). In this experiment, ShIRT was only included for comparison purposes and to demonstrate the dependence and importance on registration algorithm/parameter selection.
Large Displacement
170 A Canny edge detection algorithm [107] (built-in to MATLAB) was used on the motion images to identify an approximation of the sliding boundary position. The default lower and upper thresholds for ‘edge strength’ were used (determined by MATLAB based on the input data). The results are shown in Figure 6.4.
(a) (b)
Figure 6.4: Result of a Canny edge detection on the motion images shown in Figure 6.3 [(a) = ANTs, (b) = ShIRT]. The orange oval highlights an area where the edge detection algorithm was unable to produce a closed
loop on the result from the ShIRT registration.
The edge detection algorithm successfully identified a perimeter for the abdominal surroundings of the ANTs registration result (Figure 6.4a). The boundary identified was achieved fully automatically without the need for any user interaction and could be used to define a starting point for the ROI. The ShIRT registration did not result in a closed boundary as the edge corresponding to a portion in the lower abdomen was not detected, indicated by the orange oval in Figure 6.4b. As mentioned at the start of Chapter 2, this gap highlights a major drawback of edge detection techniques and points towards a lack of robustness.
The boundary in Figure 6.4 appears thick because the edge detection was performed on an image of the displacement at each node. A node spacing of four was used, meaning the images were a quarter of the resolution of the original images and the boundary is therefore four pixels thick. It should be noted the cine-MRI chosen was a ‘favourable’ example as large amounts of
171 motion within the abdominal contents could be observed. Less ideal images with less movement could present problems for this method.
Overlaying the result of the boundary in Figure 6.4a onto one of the cine-MRI frames produces the result in Figure 6.5.
Figure 6.5: ROI drawn manually compared to an automatic sliding boundary detection using segmentation of motion resulting from an ANTs registration. Results are overlaid on a frame of the original cine-MRI sequence.
Orange arrows highlight areas where the automatic boundary deviates from the manual boundary.
Figure 6.5 shows a good correlation between the edge detection result and the manually drawn ROI. A particularly close match is observed along the majority of the posterior portion of the sliding boundary and lower to mid abdominal wall. The areas highlighted by the orange arrows indicate apparent (but small) deviations from the manual boundary. A further problem which could be encountered is that the images of motion represent average motion over all frames. Therefore, the edge which is detected may not map to any particular frame. This could lead to the user having to modify all the calculated ROI vertex positions, resulting in no time saving over drawing the ROI manually. Overall, if similar results to those in Figure 6.5 were replicated across more patients, this initial estimate of the sliding boundary could reduce processing time.
This investigation has explored edge detection of the motion image, but given the pitfalls of edge detection, other segmentation algorithms such as intensity threshold techniques or region
172 growing may be more suitable options. Edge detection was explored as this could have potentially been easily implemented as a fully automated technique without the need for user interaction. The region of the abdominal contents in the images in Figure 6.3 do not display sharp changes in intensity; they are reasonably homogenous regions. For this reason, region growing techniques with a dynamically varying homogeneity criterion remains a possibility and may be a target for further investigation.