Abstract
Rodent models of myocardial infarction (MI) have been extensively used in biomedical research towards the implementation of novel regenerative therapies. Permanent ligation of the left anterior descending (LAD) coronary artery is a commonly used method for inducing MI both in rat and mouse. Post-mortem evaluation of the heart, particularly the MI extension assessment performed on histological sections, is a critical parameter for this experimental setting. MI extension, which is defined as the percentage of the left ventricle affected by the coronary occlusion, has to be estimated by identifying the infarcted- and the normal-tissue in each section. However, because it is a manual procedure it is time-consuming, arduous and prone to bias. Herein, we introduce semi-automatic and automatic approaches to perform segmentation which is then used to obtain the infarct extension measurement. Experimental validation is performed comparing the proposed approaches with manual annotation and a total error not exceeding 8% is reported in all cases.
Keywords: Infarct extension evaluation, image segmentation, region growing, otsu, k-means, meanshift, watershed.
Introduction
Acute myocardial infarction is a major public health problem, resulting mainly from the occlusion of coronary arteries, due to the build-up of arteriosclerotic plaques, and the establishment of tissue ischemia eventually leading to end-stage heart failure. Permanent ligation of the left anterior descending (LAD) coronary artery in animal models, including the rat and the mouse, is a commonly used method for reproducing several of the human-associated pathological events. This surgical procedure also allows the implementation of pre-clinical models of disease which are a pre-requisite for testing cell/drug-therapies before proceeding into clinical trials [1]. The tissue extension of the induced myocardial infarction, which is defined as the percentage of the left ventricle affected by coronary occlusion, is a critical parameter to evaluate the effect of any applied therapy at the experimental setting.
Fig. 1. Experimental myocardial-infarction mouse model. A – Macroscopic view of 21 days post-infarction heart; black arrow indicates the anatomical location of the LAD coronary artery ligation. B and C – Histological cross-sections of apical and mid region of LV stained with Masson Trichrome. Apex and free LV wall are fully compromised by ischemia, wich is illustrated by the collagen deposition (blue region) replacing the viable myocardium tissue (red region).
This is calculated as the average value of infarct extension over all cross-sections of the dissected heart stained with Mas-son’s Trichrome, a histological stain that enables the identification of collagen deposition, a hallmark of established infarction [1,2]. To determine the infarct extension it is necessary to indentify the infarcted-tissue (blue area) and the normal-tissue (red area) in each section (Figure 1). Currently these tasks are performed manually by the biologists, which is a time-consuming and arduous endeavor. The latter is a driving force to the development of approaches to aid the analysis of the experimental MI extension. Our approaches entail the segmentation of the cross sections of the heart, which can be performed by means of automated image processing techniques.
The multiple techniques that may be applied to the segmentation of animal tissue can be discriminated in two major classes: automatic and semi-automatic techniques. In the former case the user needs to define initial parameters for each image in order to start the segmentation. Thus, automatic segmentation requires only the validation of the initial parameters and then the algorithms segment all the images in study without further user intervention.
Region growing is a semi-automatic technique that can be used to segment the cross sections of the heart. Alattar et al. describe the use of this technique in segmentation of the left ventricle in cardiac MRI (magnetic resonance imaging) scans [3]. This technique exploits spatial context by grouping pixels or sub-regions into larger regions. Homogeneity is the main criterion for merging the regions. However, the selection of similarity criteria used depends on the problem under consideration and also on the type of image data available [4,5].
Regarding automatic segmentation there are techniques such as thresholding, region based segmentation and cluster based segmentation that can also be used in tissue segmentation [4,6]. Sharma et al. introduce the segmentation of CT (com-puturized tomography) abdomen images using a threshold segmentation technique to separate different regions in the images [6]. In a thresholding technique a single value (threshold) is used to create a binary partition of the image intensities. All intensities greater than the threshold are grouped together into one class and those below the threshold are grouped into a separate class [4,7]. Watershed is also a method applied in medical images segmentation. It is a region based segmentation which involves the concept of topography and hydrography. Hamarneh et al. present MR (magnetic resonance) cardiac images segmented with watershed transform [8]. Watershed can be described as a flooding simulation. Watersheds, or crest lines, are built when the water rise from different minima. All pixels associated with the same catchment basin are assigned to the same label [8,9]. For image segmentation, the watershed is usually, but not always, applied to a gradient image. Since real digitized images present many regional minima in their gradients, this typically results in an excessive number of catchment basins (over-segmentation) [5,9]. Ahmed et al. describe the segmentation of MR brain images using k-means clustering algorithm [10]. K-means segments the entire image into several clusters according to some measure of dissimilarity [8,10]. Mean-shift technique has also been used in segmentation of MR brain images [11]. The mean-shift algorithm is a clustering technique which does not require prior knowledge of the number of clusters, and does not constrain the shape of the clusters, requiring only the definition of the radius of the kernel used [9].
We use these techniques to (1) segment all histology processed cross-sections of the excised mouse-hearts, (2) calculate the infarct extension and finalize by (3) comparing the results with manual annotation.
This paper is organized as follows: Section 2 introduces the methodology and describes automatic and semi-automatic techniques used in segmentation of the heart, Section 3 defines how to measure the infarct extension Section 4 presents the results obtained and finally the conclusion is presented in Section 5.
Methodology
To obtain the infarct extension it is necessary to segment the different tissues in each cross section of the heart. This can be performed with semi-automatic and full automatic techniques. Within the existing semi-automatic methods for image segmentation we had chosen to use region growing due to its speed and ease of interaction. Otsu thresholding technique, watershed segmentation, k-means and mean-shift clustering are the fully automatic techniques that we selected to segment the cross sections of the heart.
In order to improve the segmentation process we applied noise reduction. For this task we tested the method BM3D [12] and the Gaussian filter [5]. The results showed that there were no significant differences in the final segmentation between images filtered with either methods. This lead us to choose the Gaussian filter since BM3D filtering is considerably slower. Noise reduction is applied to all images prior to segmentation.
Semi-automatic Tissue Segmentation
Region growing exploits spatial context by grouping pixels or sub-regions into larger regions according to some criterion. The average gray level information is the criterion chosen for merging the neighboring regions in our work. Regions are merged if they satisfy the chosen criterion and no merging occurs when the criterion is not met [5,4].
Fig. 2. Segmentation of normal tissue using the Red channel (1) and infarcted tissue using the Blue channel (3) in a cross section of the heart by region growing technique. The results of the segmentation process are binary images (2 and 4). The red points indicate the initial positions of the region growing process.
The user needs to specify the initial points to begin the segmentation process. For the task of segmenting the normal and the infarcted heart tissue it is necessary to define the initial points for the process of region growing in each of the tissue-conditions. To segment the normal-tissue we used the gray level information present in the Red channel. For the segmentation of the infarcted-tissue we used the gray level information from the Blue channel. The result is a set of binary images, one for each tissue condition (Figure 2). Results are improved using morphological operations, for example to fill small holes inside the segmentation results.
Given the segmentation areas we can then calculate the infarct extension.
Automatic Tissue Segmentation
To automatically segment the different tissue-conditions in each of the heart cross-sections we use otsu thresholding, watershed segmentation, k-means and mean-shift clustering. All these image segmentation techniques allow the partition of the image in regions which we can associate to the distinct tissue-conditions by analyzing their color.
Otsu thresholding technique selects an adequate threshold of gray level for extracting objects from their background. This threshold is used to convert an intensity image to a binary one. All intensities greater than the threshold are grouped together into one class and those below the threshold are grouped into a separate class [13].
Using this technique, with different channels of the RGB image, we can obtain segmentations of the normal and infarcted-tissue. High values of image intensities in the Red channel relate to normal-tissue. The Blue channel has high image intensity values in infarcted areas. Based on these relationships between the Red and Blue color channels and the tissue properties we decided to subtract the Blue channel to the Red channel for the segmentation of the normal-tissue. We subtract the Red channel to the Blue channel for the segmentation of the infarcted-tissue (Figure 3 (a)).
Watershed technique is based on immersion simulation. The input image is considered as a topographic surface which is flooded by water starting from regional minima. Watershed lines are formed on the meeting points of water coming from distinct minima. All pixels associated with the same catchment basin are assigned to the same label [14,15].
Fig. 3. Segmentations of a heart cross section and identification of the normal and infarted tissue: (a) Combination of the channels (1 and 2) and respective otsu thresholding results (3 and 4), (b) Watershed segmentation result, (c) K-means and (d) Mean-shift clustering results (top) and respectively identification of the regions.
For the application of this technique we use the same image channel combination as for otsu thresholding. Performing watershed segmentation originates an oversegmentation of the tissue since it has many regional minima. However, by comparing the color intensities in each region we are able to decide if each region is from normal-tissue or from infarcted-tissue. Using also the otsu thresholding technique that allows to easily obtain the full tissue segmentation we focus our analysis only on the tissue region. The resulting tissue areas are coherent with normal/infarct tissue -areas (Figure 3 (b)).
K-means clustering technique assigns cluster labels to data points from the entire image [8]. For this technique we use the information of the three channels, selecting three clusters which will correspond to the background, normal-and infarcted-tissue. After obtaining the segmentation result we identify each segmented cluster from its average color intensity. To improve the segmentation we fill the holes using morphological operations (Figure 3 (c)).
Mean-shift clustering technique does not require prior knowledge of the number of clusters and only needs the definition of the radius of the kernel used. As in the previous technique we decided to obtain at most three clusters. If we obtain more than three clusters we iteratively increase the radius of the kernel used (Figure 3 (d)).
In this case we base our segmentation on the Red and Blue channels since they lead to better results than the use of all channels.
Following the segmentation results of full automatic and semi-automatic techniques we can measure the infarct extension.
Infarct Extension Evaluation
To better understand the calculation of the infarct extension we must analyze the different regions in the heart. In Figure 4 (A) we can observe the heart bounded by the exterior black continuous line.
Fig. 4. Image of a heart cross section. A – The heart is bounded by the outside black continuous line which includes the left ventricle (LV) and right ventricle (RV) separated by line a. The interior black continuous line is identifying the lumen of the left ventricle and the region marked by lines shows the tissue with infarct. The dotted line is the midline between the inside and outside black continuous lines. B – Scheme of a cross section of the heart.
It is formed by right and left ventricles. The first consists only in normal-tissue. The left ventricle includes the infarcted-tissue, which is represented by the shaded region, lumen, which is bounded by the interior black continuous line and also normal-tissue. The infarct extension is usually calculated by two different methods:
Area measurement - Infarct extension is calculated by dividing the infarct area by the area of the heart tissue [1,2] (Figure 4 (A)). This is trivial based on the segmentation results obtained.
Midline length measurement – Infarct extension is calculated by dividing the midline infarct length by the length of midline [2] (Figure 4 (A)). Figure 4 (B) shows a scheme that represents a cross-section of the heart. To perform the midline measurement we first automatically find the midline by tracing lines from the centre of the lumen to the outside of the tissue. The midline is given by the middle distance between tissue borders. The points of the middle line where there is infarcted-tissue in bigger percentage than the normal-tissue (in the radial direction) are considered infarcted points. Secondly we divide the length of infarct midline by the length of the midline. To obtain the lumen of the heart we get a segmentation of all the heart tissue by otsu thresholding technique, which is trivial and we identify the biggest hole inside that segmentation.
The infarct extension is defined as the mean value of infarct extension in all the cross-sections of the heart.
Infarct extension is evaluated in the heart cross-sections considering or not the right ventricle [1,2] (Figure 4 (A)). However, it is not easy to find a robust way to remove the RV as it is variable in morphology and most biologists vary in their assessment of where the RV ends and the LV begins. As such, we perform both the analysis on the full heart, including the RV, and also obtain results on the LV only by removing the RV through image editing tools (manually).
Table 1. Results of infarct extension measurement in mice’s hearts. The results are the average value of infarct extension obtained in transverse sections of each mouse heart. In the manual analysis the results only consider the left ventricle.
|
Infarct extension |
||||||
|
Midline length measurement |
||||||
|
Heart |
Manual |
Region growing |
Otsu |
Watershed |
K-means |
Mean-shift |
|
#1 with RV |
- |
39% |
38% |
39% |
37% |
34% |
|
#1 without RV |
43% |
38% |
40% |
41% |
38% |
36% |
|
#2 with RV |
- |
47% |
47% |
48% |
46% |
47% |
|
#2 without RV |
52% |
48% |
48% |
49% |
47% |
49% |
|
Area measurement |
||||||
|
Heart |
Manual |
Region growing |
Otsu |
Watershed |
K-means |
Mean-shift |
|
#1 with RV |
- |
29% |
18% |
17% |
17% |
14% |
|
#1 without RV |
22% |
25% |
21% |
21% |
20% |
17% |
|
#2 with RV |
- |
34% |
29% |
31% |
28% |
33% |
|
#2 without RV |
36% |
38% |
32% |
36% |
32% |
36% |
Results
The infarct extension was calculated manually and automatically in two independent hearts. The calculation was performed both on the whole cross-section tissue and also without considering the right heart ventricle for comparison. To automatically segment the tissue without taking into account the right ventricle we manually remove this region before the segmentation process. Table 1 shows the results for the infarct extension evaluation using our approaches and the manual annotation. The results are the average value of the infarct extension over all cross-sections of each independent heart.
Differences between the proposed approaches and manual annotation are never greater than 8% in the case of the evaluation considering the right ventricle. Removing the right ventricle the differences are never greater than 7%. The differences among the proposed approaches in mice’s hearts considering the right ventricle tissue are at most 15% and without this are never greater than 8%.
Conclusion
The proposed approach enabled the full and semi-automatic calculation of infarct extension. The results obtained using our approaches were in close agreement with the manual annotation with differences never higher than 8%. The segmentation allowed an analysis of the infarct extension in a fraction of the manual method measure time.
Within the automatic segmentation approaches, the watershed technique produced better results, with the differences never above 5% (reduced to 3% by removing the right ventricle). The differences from the semi-automatic approach used were at most 7% considering the right ventricle (5% without this one). Although the differences were slightly higher in the semi-automatic approach, the biologists prefer the possibility to control the segmentation results in relation to fully automatic approaches.
Future research will focus on integrating automatic image segmentation methods with anatomical models. This will enable the automatic segmentation and measurement of only the left ventricle of the heart, leading to better results.
