Automatic segmentation of myocardium at risk from contrast enhanced SSFP CMR: validation against expert readers and SPECT

Abstract

Background: Efficacy of reperfusion therapy can be assessed as myocardial salvage index (MSI) by determining the size of myocardium at risk (MaR) and myocardial infarction (MI), (MSI = 1-MI/MaR). Cardiovascular magnetic resonance (CMR) can be used to assess MI by late gadolinium enhancement (LGE) and MaR by either T2-weighted imaging or contrast enhanced SSFP (CE-SSFP). Automatic segmentation algorithms have been developed and validated for MI by LGE as well as for MaR by T2-weighted imaging. There are, however, no algorithms available for CE-SSFP. Therefore, the aim of this study was to develop and validate automatic segmentation of MaR in CE-SSFP.

Methods: The automatic algorithm applies surface coil intensity correction and classifies myocardial intensities by Expectation Maximization to define a MaR region based on a priori regional criteria, and infarct region from LGE. Automatic segmentation was validated against manual delineation by expert readers in 183 patients with reperfused acute MI from two multi-center randomized clinical trials (RCT) (CHILL-MI and MITOCARE) and against myocardial perfusion SPECT in an additional set (n = 16). Endocardial and epicardial borders were manually delineated at end-diastole and end-systole. Manual delineation of MaR was used as reference and inter-observer variability was assessed for both manual delineation and automatic segmentation of MaR in a subset of patients (n = 15). MaR was expressed as percent of left ventricular mass (%LVM) and analyzed by bias (mean ± standard deviation). Regional agreement was analyzed by Dice Similarity Coefficient (DSC) (mean ± standard deviation).

Results: MaR assessed by manual and automatic segmentation were 36 ± 10% and 37 ± 11%LVM respectively with bias 1 ± 6%LVM and regional agreement DSC 0.85 ± 0.08 (n = 183). MaR assessed by SPECT and CE-SSFP automatic segmentation were 27 ± 10%LVM and 29 ± 7%LVM respectively with bias 2 ± 7%LVM. Inter-observer variability was 0 ± 3%LVM for manual delineation and -1 ± 2%LVM for automatic segmentation.

Conclusions: Automatic segmentation of MaR in CE-SSFP was validated against manual delineation in multi-center, multi-vendor studies with low bias and high regional agreement. Bias and variability was similar to inter-observer variability of manual delineation and inter-observer variability was decreased by automatic segmentation. Thus, the proposed automatic segmentation can be used to reduce subjectivity in quantification of MaR in RCT.

Clinical trial registration: NCT01379261. NCT01374321.

Figures

Fig. 1

Automatic segmentation algorithm. The new automatic algorithm for segmentation of myocardium at risk (MaR) in CE-SSFP lets the user define the culprit artery and the rotation of the left ventricle as input. The algorithm consists of four processing blocks, surface coil intensity correction, intensity classification by Expectation Maximization (EM), segmentation based on a priori information on MaR and incorporation of infarct region from LGE images

Fig. 2

Correlation and bias for automatic segmentation and threshold methods against manual delienation. Correlation of MaR as % of LVM (left column) and Bland-Altman plot of MaR bias as % of LVM (right column) for the automatic segmentation algorithm (first row), threshold of 2SD from remote (second row), FWHM (third row) and Otsu (fourth row), all against manual delineation. The line of identity is shown as a solid line for all correlations plots and mean bias (solid line) and mean ± two standard deviations (dashed line) is shown for all Bland-Altman plots

Fig. 3

Example of automatic segmentation and manual delineation of MaR in CE-SSFP. Typical MaR segmentation in all left ventricular short axis slice images from one patient in end-diastole (ED, top panel) and end systole(ES, bottom panel), for automatic segmentation by Segment MaR CE-SSFP, shown in white, and manual delineation, shown in purple. Endocardial borders are shown in red and epicardial border in green. For this patient MaR by manual segmentation was 44 %LVM and by automatic Segment MaR CE-SSFP 43 % LVM with a regional agreement DSC of 0.85

Fig. 4

Correlation and bias against SPECT for automatic segmentation and manual delineation in CE-SSFP. Correlation of MaR as % of LVM (left column) and Bland-Altman plot of MaR bias as % of LVM (right column) against SPECT for automatic segmentation algorithm Segment MaR CE-SSFP (top row) and manual reference delineation (bottom row). The line of identity is shown as a solid line for all correlations plots and mean bias (solid line) and mean ± two standard deviations (dashed line) is shown for all Bland-Altman plots. Correlation and Bland-Altman plots for manual delineation in CE-SSFP against SPECT (bottom row) are adopted from Sorenson et al. [4]

Fig. 5

Analysis of incremental value of blocks in the automatic segmentation algorithm. Incremental value of each block in the automatic segmentation algorithm analyzed by bias to manual delineation as %LVM, left panel and by regional agreement as Dice similarity coefficient DSC (right panel). Bias and DSC was calculated with segmentation based on only intensity classification by Expectation Maximization and calculated after the addition of the processing blocks of intensity correction, a priori on myocardium at risk (MaR) and infarct region from late gadolinium enhancement (LGE). For each block of the algorithm the upper limit of the box indicate upper quartile, middle line indicate median, lower limit of box indicate lower quartile, whiskers indicate minimum and maximum points within 1.5 interquartile range and points (+) indicate outliers. Bias zero is shown as dotted black line in the left panel, DSC above of 0.7 indicates good regional agreement [30], and is shown as dotted black line in the right panel. Two sided paired t-test was performed for each block in comparison to previous block and first block, ns: non significant, ***: p < 0.0001

Fig. 6

Model of maximal extent for perfusion territory of each culprit artery. Bulls-eye representation of maximal extent model for the perfusion territories of left anterior descending artery (LAD), left circumflex artery (LCx), right coronary artery (RCA), and left main artery (LM). Models for LAD, LCX and RCA were defined in consensus by three experienced observers in an extended 17- segment AHA model and models for LM were defined from the models of LAD, LCX and RCA. The 17-segment model is extended to three slices in each of the basal, mid-ventricular and apical zones and 24 sectors in each slice. Black sectors are included in the maximal extent model. The septal part of the left ventricle is represented in the left of the bulls-eye plot, the lateral part in the right, anterior part in the top, inferior part in the bottom, the apical slices in the center and the basal slices in the outer part of the bulls-eye plot

Fig. 7

Surface coil intensity correction. Histogram of myocardial intensities within the myocardium before intensity correction (top panel) and after intensity correction (bottom panel), represented in black for culprit region and white for remote region. After intensity correction a decreased standard deviation of remote region is obtained and the mean intensity of the culprit region is higher than remote region. Myocardial intensities as after intensity correction (bottom panel) are used as input to the EM-algorithm with remote region and culprit region as initial classification and dashed blue lines indicating the constraints. Intensities below the 50th percentile of the remote region are kept classified as normal myocardium through the iterations of the EM-algorithm (lower dashed blue line). Intensities above the 75th percentile of the intensities within the culprit region are kept classified as MaR through the iterations of the EM-algorithm (upper dashed blue line). Dashed red line indicates the resulting cut off 0.5 of the Bayesian probability of MaR resulting from the EM-algorithm after intensity correction

Fig. 8

Utilization of a priori information on extent and transmurality. Three short axis slice, basal, mid and apical with MaR segmentation, shown in white, based on only the Bayesian probability of MaR (top row) and with the addition of a priori regional criteria on extent and transmurality (bottom row)

Fig. 9

Utilization of infarct segmentation from LGE. Three short axis slices, basal, midventricular and apcial, from LGE (top row) with infarct region in yellow, infarct core in pink and microvascular obstruction in red, and CE-SSFP (bottom row) with automatic segmentation by Segment MaR CE-SSFP after utilization of information on infarct region from LGE. The large region of hypoenhancement in CE-SSFP is included as MaR by utilizing the segmentation of infarct region