We have developed and tested a highly automatic method for estimating the extent of myocardial oedema by using bright blood T2-weighted CMR in patients with acute myocardial infarction. Physician judgment is retained as the initial step in image analysis and following this initial evaluation, images are subjected to automatic segmentation of the LV wall and oedema territory. High levels of agreement for the proposed approach are found when compared to a standardized and validated manual approach for myocardial oedema analysis on T2-weighted CMR (Dice similarity coefficient >0.7). Furthermore our new approach is also seen to outperform two other computerized methods [17, 21]. Although the results from Sjogren et al.  have a higher Dice similarity coefficient than our approach, their results were obtained from manually segmented LV boundaries. Recognising the value of clinical judgement, we retained user input as a key first step in the process. Thus, observer judgement is initially required to assess for the presence and distribution of hyperintense myocardial regions and place these observations in a clinical context (e.g. acute myocardial infarction). Thus clinical judgement is the first step in image interpretation prior to automatic analyses, which is why the proposed method is considered as highly automatic rather than fully automatic.
LV wall boundary segmentation (semi-/fully automatic) with CMR has been extensively studied [35–37]. However accurate LV segmentation remains a challenge especially for quantitative analysis of global and regional cardiac function. In the current study, a level set method without re-initialization was used for LV wall segmentation. The comparison with manual segmentation shows that this method can generate good LV wall boundaries for further oedema quantification. Unlike previous studies [16, 21, 22], in which the computerized method for oedema delineation required manual defined LV wall boundaries (time consuming, possible large inter/intra-observer variations), our approach directly links the oedema delineation to the automatic LV wall segmentation procedure. However the uncertainties arising from automatic LV wall segmentation will contribute to the final extent of oedema, which in turn will contribute to errors in oedema quantification. As in Figure 3 and Figure 7, false positive pixels may be included, illustrating the importance of accurate LV boundary segmentation for automatic oedema quantification. If the manual approach to delineation of endo- and epicardial boundaries is used then the Dice similarity for oedema estimation is 0.8±0.05, higher than using automatically delineated LV boundaries, also close to the value in the inter-observer study.
Gudbjartsson et al.  suggested that for magnitude MR images, if signal to noise ratio, (SNR: defined as A/σ, A is the mean pixel intensity in the absence of noise, and σ denotes the standard deviation of the noise), is less than 1, the distribution of MR image intensity can be approximated by Rayleigh distribution; while when SNR as small as A/σ=3, it can be approximated by Gaussian distribution. Four random patients were selected for evaluating SNR. σ was averaged from the original MR images measured from regions in background air. After windowing process, σ was assumed to be the same, the averaged pixel intensity M in oedema regions was 103, and 3.85 in unaffected regions. Then SNR for oedema regions was
, larger than 3, for unaffected regions: A/σ=0.98. The analysis indicates Rayleigh-Gaussian distribution is applicable in this study. Furthermore, the comparison with Kadir’s method  shows that our approach performs better for oedema quantification even though the results from Kadir’s work produced acceptable Dice similarity scores. Johnstone et al  used a Gaussian mixture model to fit the myocardium intensity histogram with manual delineated LV boundaries and the agreement using Dice similarity was low (0.5). Since Kadir’s work and Johnstone’s work are based on the assumption that myocardium intensity has a Gaussian distribution, the higher agreement in our study also may suggest that the Rayleigh-Gaussian distribution is a good approximation for bright-blood T2-weighted oedema images. Our use of the Rayleigh-Gaussian distribution approximation on bright blood T2-weighted oedema images is novel and more studies are needed to further evaluate the performance and applicability of this approach.
By incorporating a prior model of the maximal extent of user defined culprit, Sjogren et al  improved the performance of Johnstone’s method for oedema segmentation from base to apex. However Sjogren’s method still required manual delineation of endo/epi-cardial boundaries and the oedema region was detected based on pre-defined regional analysis rather than the pixel-wise method used widely in other studies. Our method provides comparable results with similar bias of -0.4±4% with automatically detected LV boundaries, but lower Dice similarity coefficient 0.74±0.06. The lower Dice similarity coefficient in our study could be due to the pixel wise analysis for initial oedema segmentation rather than the regional analysis and the less accurate automatically delineated LV boundaries compared to manually segmented LV boundaries. Other factors may also be relevant such as variation in pathology between subjects (e.g. myocardial haemorrhage within the infarct zone), and image artefacts. In fact, some prior information is included in our method by retaining the option to alter the grey-scale window and level and observer judgement on the presence of oedema. Considering the superior diagnostic performance of bright blood T2-weighted oedema imaging over other methods (e.g. dark blood STIR MRI) , we still think observer input is important to avoid automated delineation of artefacts. Currently most computerized methods for oedema quantification do not take advantage of the information from image windowing [21, 22], which we believe facilitates image assessment as illustrated in Figure 2(b).
After thresholding, alternate morphological filtering was applied to produce smoothed oedema regions without changing the overall shape. A kernel size of 2 pixels was used in the beginning for the closing and opening process followed by a kernel size of 5 pixels for the closing process. The kernel size of 5 pixels for the last closing process is in line with the approach by Hsu et al. (5 mm) . The morphological filtering might still not be able to close out a signal void within the oedema region, as in Figure 3(d). Accordingly an algorithm for detecting dark zones, which potentially may be myocardial haemorrhage, is desired. In Johnstone’s study , a hyperintense region of possible interest but <1 g mass was considered to be noise and excluded from oedema quantification. In our dataset, since each oedema region is approximately perpendicular to the long axis of the LV and the slice interval is 10 mm, if mid-ventricular dimensions are adopted (cavity diameter: 45 mm, wall thickness: 10 mm), then the average oedema mass at each slice is approximately 6 g for manual delineation with an assumed density of 1.05 g cm3. In oedema region feature analysis in our study, approximately one-fifth of the main oedema area was considered to be a critical threshold for discrimination from noise. This threshold area corresponds to an average mass of 1.2 g. To improve the post-processing, minimum distance constraint criteria with the main oedema region was applied. By increasing the degree to 20 in the third step of feature analysis, and in the second step, the degree was changed to 40 at the same time. The result showed that the oedema extent was 27±8% with a mean Dsc of 0.73±0.06 related to the manual quantification. This analysis indicates that the oedema extent may not be sensitive to the choice of the minimum distance constraint. If no any feature analysis was applied, then the mean oedema extent was 30±8% (Dsc: 0.7±0.06), suggesting that oedema feature analysis is essential. Another potential source of error for overestimation of the oedema area is incorrect placement of the endocardial border within the bright LV cavity blood pool. In future a dark blood T2-prepared CMR method with steady state free precession readout might overcome this problem.
In this study, we intended to minimize user-inputs for oedema quantification, therefore there is no manual correction for automatic LV boundary segmentations and oedema delineations. As shown in Figure 3 and Figure 7, false positive and positive false pixels are present for reasons including (1) inaccurate automatic LV boundary segmentation, and (2) large dark areas inside the oedema regions. We believe that with necessary manual correction, the accuracy of oedema quantification will be highly improved.
In order to use our method attention is needed for: (1) the windowing process, which could introduce variability due to the observer’s experience. Our sensitivity study of the windowing process shows that the automatic approach is able to quantify oedema region consistently if the observer is trained properly on T2-weighted oedema images. (2) The definition of optimal threshold value, which might be different from 0.7 when actual oedematous regions are available for comparison, thus it needs to be carefully selected for different studies. (3) The oedema region feature analysis, which may not be applied to patients such as with myocarditis. Other limitations: (1) the manual input in our method involves user-defined adjustment of the greyscale level. Potentially, future technical developments may enable this step to be removed. Secondly there is no automatic decision making step for the presence of oedema. However, we suggest this aspect should be considered a strength of the method since physicians’ judgment is still retained; (2) Improvement in methods for automatic LV boundary segmentation and dark zones detection (such as haemorrhage) within the oedema regions are needed for more accurate oedema quantification; (3) 3D anatomical structure information of oedema, culprit coronary artery assignment  and myocardial infarction could be integrated in order to optimize the post-processing and segmentation results; (4) our method has been tested in a reasonably large cohort of patients with acute myocardial infarction. The performance of our method in other pathological conditions, such as acute myocarditis, needs further evaluation. Future studies should validate the method in animal experiments and in a larger cohort of acute MI patients.