Introduction
-
We review both the segmentation techniques and the structural and functional indices that use the extracted boundaries in their calculations. This will provide a better understanding of the role of cardiac segmentation to the final clinical assessment.
-
We also list the quantitative evaluations for the accuracy of both the segmentation and functional analysis to provide the readers with an overview of the level of performance of the existing techniques so far.
-
We provide a more comprehensive review of segmentation techniques for all cardiac chambers, including the left ventricle (LV), the right ventricle (RV), the left atrium (LA), and the whole heart.
-
We also include the use of long-axis images in cardiac segmentation as they play an important role in clinical use of CMR.
Abbr. | Structure | Calculation methods | Requirement and parameters | Exemplar applications | Normal range | |
---|---|---|---|---|---|---|
Left ventricle
| ||||||
End-diastolic volume | LVEDV | ± Papillary muscles + Outflow tract | Single area-length method Bi-plane area-length method Simpson’s method Direct measurement | 2-Chamber LAX view and axis length 2-Chamber, 4-chamber LAX view and axis length Cross-sectional area on each SAX slice and slice thickness A number of contiguous voxels | Dilated cardiomyopathy | M: 156 ± 21 mL, F: 128 ± 21 mL [33] M: 160 ± 29 mL, F: 135 ± 26 mL [34] |
End-systolic volume | LVESV | ± Papillary muscles + Outflow tract | Single area-length method Bi-plane area-length method Simpson’s method Direct measurement | 2-Chamber LAX view and axis length 2-Chamber, 4-chamber LAX view and axis length Cross-sectional area on each SAX slice and slice thickness A number of contiguous voxels | Dilated cardiomyopathy | M: 53 ± 11 mL, F: 42 ± 9.5 mL [33] M: 50 ± 16 mL, F: 42 ± 12 mL [34] |
Myocardial mass | LVM | ± Papillary muscles and trabecular tissue | (LVVepi − LVVendo) × 1.05 | LVVepi: Left Ventricle Epicardial Volume LVVendo: Left Ventricle Endocardial Volume | Hypertension, hypertrophic cardiomyopathy | M: 146 ± 20 g, F: 108 ± 18 g [33] M: 123 ± 21 g, F: 96 ± 27 g [34] |
Stoke volume | LVSV | LVEDV − LVESV | End-diastolic and end-systolic volumes | Aortic insufficiency, aortic stenosis | M: 104 ± 14 mL, F: 86 ± 14 mL [33] M: 112 ± 19 mL, F: 91 ± 17 mL [34] | |
Ejection fraction | LVEF | (LVEDV − LVESV)/LVEDV × 100 % | Stroke volume and end-diastolic volume | Heart failure, hypertrophic cardiomyopathy | M: 67 ± 4.5 %, F: 67 ± 4.6 % [33] M: 69 ± 6.0 %, F: 69 ± 6.0 % [34] | |
Cardiac output | LVCO | LVCO = LVSV × HR | Stroke volume and heartbeat rate | Hypertension, congestive heart failure | 4–8 L/mina
| |
Wall thickness | – | − Papillary muscles − Trabecular tissue | Radial method Centreline method | Endocardial and epicardial contours on short-axis image slices, a centre point Endocardial and epicardial contours on short-axis image slices, a centreline and anatomical reference points | Myocardial infarction, hypertension, hypertrophic cardiomyopathy | M: Basel: 7.8 ± 1.1 mm; Mid: 6.3 ± 1.1 mm; Apical: 6.4 ± 1.1 mm, F: Basel: 6.4 ± 0.9 mm; Mid: 5.3 ± 0.9 mm; Apical: 5.9 ± 0.9 mm [35] |
Wall thickening | – | (wall thicknessed − wall thicknesses)/wall thicknessed × 100 % | Average end-systolic wall thickness and average end-diastolic wall thickness | Basal: 73 ± 31 % Mid: 79 ± 26 % Apical: 64 ± 30 % [36] | ||
Strain analysis | LSA | Global coordinates Local coordinates | Lagrangian or Eulerian strain rate Longitudinal, radial, and circumferential strain | Initial location and deformed location Strain tensor | Myocardial infarction, ischemia, and ventricular dyssynchrony | |
Right ventricle
| ||||||
End-diastolic volume | RVEDV | ± Papillary muscles + Trabecular tissue | Simpson’s method | Cross-sectional area on each slice and slice thickness | Arrhythmogenic right ventricular cardiomyopathy, congenital heart diseases | M: 190 ± 33 mL, F:148 ± 35 mL [34] M:163 ± 25 mL, F: 126 ± 21 mL [37] |
End-systolic volume | RVESV | ± Papillary muscles + Trabecular tissue | Simpson’s method | Cross-sectional area on each slice and slice thickness | Arrhythmogenic right ventricular cardiomyopathy | M: 78 ± 20 mL, F: 56 ± 18 mL [34] M: 57 ± 15 mL, F: 43 ± 13 mL [37] |
Stroke volume | RVSV | RVEDV − RVESV | End-diastolic and end-systolic volume | Pulmonary arterial hypertension | M: 113 ± 19 mL, F: 90 ± 19 mL [34] M: 106 ± 17 mL, F: 83 ± 13 mL [37] | |
Ejection fraction | RVEF | RVSV/RVEDV × 100 % | Epicardial and endocardial volume | Pulmonary arterial hypertension, congestive heart failure | M: 59 ± 6.0 %, F: 63 ± 5.0 % [34] M: 66 ± 6.0 %, F: 66 ± 6.0 % [37] | |
Cardiac output | RVCO | RVCO = RVSV × HR | Stroke volume and heartbeat rate | Ventricle failure with cardiomyopathy, pulmonary arterial hypertension | 5.25 L/minb
| |
Left atrium
| ||||||
Maximum volume | LAVmax
| − Confluence of the pulmonary veins and LA appendage | Single area-length method Bi-plane area-length method Simpson’s method Direct measurement Ellipse method | 2-Chamber LAX view and axis length 2-Chamber, 4-chamber LAX view and axis length Cross-sectional area on each SAX slice and slice thickness A number of contiguous voxels Longitudinal diameter, transverse diameter, and antero-posterior diameter | Atrial fibrillation, congestive heart failure, mitral valve disease | M: 103 ± 30 mL, F:89 ± 21 mL [34] |
Minimum volume | LAVmin
| − Confluence of the pulmonary veins and LA appendage | Single area-length method Bi-plane area-length method Simpson’s method Direct measurement Ellipse method | 2-Chamber LAX view and axis length 2-Chamber, 4-chamber LAX view and axis length Cross-sectional area on each SAX slice and slice thickness A number of contiguous voxels Longitudinal diameter, transverse diameter, and antero-posterior diameter | Atrial fibrillation, congestive heart failure, mitral valve disease | M: 46 ± 14 mL, F: 41 ± 11 mL [34] |
Total emptying volume (reservoir) | LAEV | − Confluence of the pulmonary veins and LA appendage | LAVmax − LAVmin
| LAVmax: LA volumes assessed at LV end-systole LAVmin: LA volumes assessed at late LV end-diastole after LA contraction | Atrial fibrillation, atrial flutter, mitral stenosis, mitral regurgitation, diastolic dysfunction, dilated cardiomyopathy, diabetes mellitus, hypertrophic cardiomyopathy, amyloidosis, and hypertension | |
Total emptying fraction (reservoir) | LAEF | − Confluence of the pulmonary veins and LA appendage | (LAVmax − LAVmin)/LAVmax × 100 % | LAVmax: LA volumes assessed at LV end-systole LAVmin: LA volumes assessed at late LV end-diastole after LA contraction | 38 ± 8 % [38] | |
Passive emptying volume (conduit) | LAPEV | − Confluence of the pulmonary veins and LA appendage | LAVmax − LAVpre A
| LAVmax: LA volumes assessed at LV end-systole LAVpre A: LA volumes assessed at LV diastole just before LA contraction | Atrial fibrillation, atrial flutter, diastolic dysfunction and diabetes mellitus | |
Passive emptying fraction (conduit) | LAPEF | − Confluence of the pulmonary veins and LA appendage | (LAVmax − LAVpre A)/LAVmax × 100 % | LAVmax: LA volumes assessed at LV end-systole LAVpre A: LA volumes assessed at LV diastole just before LA contraction | 36 ± 11 % [38] | |
Conduit volume | LACV | − Confluence of the pulmonary veins and LA appendage | LSV − (LAVmax − LAVmin) | LSV: LV stoke volume LAVmax: LA volumes assessed at LV end-systole LAVmin: LA volumes assessed at late LV end-diastole after contraction | 41 ± 14 mL [39] | |
Active emptying volume (pump) | LAAEV | − Confluence of the pulmonary veins and LA appendage | LAVpre A − LAVmin
| LAVpre A: LA volumes assessed at LV diastole just before LA contraction LAVmin: LA volumes assessed at late LV end-diastole after LA contraction | Atrial fibrillation, atrial flutter, diastolic dysfunction, dilated cardiomyopathy and diabetes mellitus | |
Active emptying fraction (pump) | LAAEF | − Confluence of the pulmonary veins and LA appendage | (LAVpre A − LAVmin)/LAVpre A × 100 % | LAVpre A: LA volumes assessed at LV diastole just before LA contraction LAVmin: LA volumes assessed at late LV end-diastole after LA contraction | 26 ± 3 % [38] | |
Right atrium
| ||||||
Maximum volume | RAVmax
| 3.08 × A2C + 3.36 × A4C − 44.4 Single plane area-length method Bi-plane area-length method | A2C is the area in 2-chamber LAX view and A4C is the area in 4-chamber LAX view 2-Chamber LAX view and axis length 2-Chamber, 4-chamber LAX view and axis length | Chronic heart failure, pulmonary arterial hypertension, tricuspid valve disease, atrial septal defect | M: 109 ± 20 mL, F: 91 ± 20 mL [40] |
References | Mode | Dim | Fundamental principles | User interaction | Test cases | Training sets | Materials | Functional analysis performance | Accuracy [distance (mm) and similarity] |
---|---|---|---|---|---|---|---|---|---|
Mitchell et al. [119] | Cine | 3D | 3D AAM | Manual segmentation on training sets | 56(18) subjects | Leave 1 out | Multiphase, SAX + LAX | LVVepi: 0.97, 0.91, 12.1 (CC&LRC) LVVendo: 0.94, 0.88, 8.4 (CC&LRC) LVM: 0.82, 0.80, 17.9 (CC&LRC) | P2S: epi: 2.63 ± 0.76, endo: 2.75 ± 0.86 |
Paragios [98] | Cine | 2D | GVF-based level-sets and contour propagation | – | A few sequences | – | Multiphase, SAX | – | – |
Stalidis et al. [95] | Cine | 3D + T | Deformable surface modelling + neural network classification | Indicate rough position of cavity and reference samples | 3(3) 2D + 1 3D datasets | Guided by user | Multiphase, SAX + LAX | – | – |
Santarelli et al. [100] | Cine + P | 2D | GVF snake | Draw rough contour of the internal cavity | 9 patients (907 images) | – | Multiphase, SAX | – | – |
Kaus et al. [104] | Cine | 3D | Prior (coupled meshes + PDM) + deformable model | Manual segmentation on training sets | 121 subjects | Leave 1 out | ED + ES, SAX | – | epi: 2.92 ± 1.38 (ES), 2.62 ± 0.75 (ED) endo: 2.76 ± 1.02 (ES), 2.28 ± 0.93 (ED) |
Yeh et al. [132] | Cine | 2D | DP-based border detection | Place region of interest which includes the whole LV | 1(0) subjects | – | Multiphase, SAX | – | – |
Gotardo et al. [106] | Cine | 2D + T | Fourier shape constraints + deformable model tracking | Specify 4 points on the desired boundary in one image | 33(33) subjects |
Leave 1 out for classifier
| Multiphase, SAX | – | – |
Jolly [92] | Cine | 2D | LV localisation + EM based classification + active contours | Crop the image to limit the localisation search space | 29 patients (482 images) | – | ED + ES, SAX | – | – |
Pednekar et al. [94] | Cine | 2D + T | Motion-map and EM guided localisation + DP-based walls extraction | – | 14 subjects | – | Multiphase, SAX | LVESV error: −10.90 mL, LVEDV error: −0.17 mL LVEF error: 7.21 % | – |
van Assen et al. [120] | Cine | 3D | 3D ASM + fuzzy inference | Manual segmentation on the basal and apical slices | 15(0) subjects + 5(5) patients | Pre-constructed atlas | –, SAX/LAX | LVVepi: 219.3 ± 41.3 mL (SAX), 243.0 ± 35.0 mL (RAD), 229.1 ± 41.6 mL (MV) LVVendo: 122.0 ± 27.3 mL (SAX), 132.5 ± 18.5 mL (RAD), 127.5 ± 28.1 mL (MV) | epi: 2.23 ± 0.46 (SAX), 2.83 ± 0.78 (RAD), 2.29 ± 0.53 (MV) endo: 1.97 ± 0.54 (SAX), 2.24 ± 0.54 (RAD), 2.02 ± 0.93 (MV) |
Lynch et al. [105] | Cine | 2D | Region-based coupled level-set | Manual insertion of a seed point | 4 slices | – | –, SAX | – | P2C: endo: 0.477 ± 0.683, epi: 1.149 ± 1.157 |
Lekadir et al. [121] | Cine | 3D | 3D ASM + outlier correction | Manual segmentation on training sets | 36 subjects | Leave 1 out | –, SAX | – | epi: 1.11 ± 0.46, endo: 0.78 ± 0.21 |
Andreopoulos and Tsotsos [122] | Cine | 3D + T | 3D AAM + hierarchical 2D ASM with temporal constraints | Indicate a few endocardial and myocardial regions | 33 subjects (7980 slices) | Threefold cross validation | Multiphase, SAX | LVVepi: 0.97, 0.98, 2.7 (CC&LRC) LVVendo: 0.95, 0.92, 4.6 (CC&LRC) | – |
Chen et al. [110] | DENSE | 2D | Optimal boundary initialisation + deformable model | Crop the input image to put LV centroids in the centre | 5(0) healthy | – | Multiphase, SAX | – | – |
Codella et al. [85] | Cine | 2D | Region growth + seeds propagation | Choose the mid-ventricular slice | 38(20) subjects | – | Multiphase, SAX | LVESV error: −1.9 ± 6.1 mL LVEDV error: 0.8 ± 5.1 mL LVV: 0.99, 0.98, −2.9 (CC&LRC) LVEF error: 3 ± 7.5 % LVEF: 0.95, 1.01, −1.9 (CC&LRC) | – |
Folkesson et al. [96] | LE | 2D | Geodesic active region + statistical KNN classifier | Manual segmentation on training sets | 4 patients (30 slices) | 7 patients (57 slices) | –, SAX | – | 1.44 ± 0.54 (P2C); 0.79 ± 0.07 (DC) |
Huang and Metaxas [111] | TA | 2D | Deformable shape and appearance model | Specify the centroid and the radius by 2 clicks | – | – | – | – | – |
Lynch et al. [108] | Cine | 3D + T | Level-set + temporal prior + EM optimised fitting | – | 6 subjects | A set of boundaries | Multiphase, SAX | – | endo: 1.25 ± 1.33 (P2C) |
Sun et al. [109] | Cine | 3D + T | Level-set + recursive estimation using temporal learning | Manual segmentation on training sets | 26 patients (234 cycles, 4680 slices) | 5 patients (42 cycles, 840 slices) | Multiphase, SAX | endo: 0.93 (DC) | |
van Assen et al. [123] | Cine | 3D | 3D ASM + fuzzy inference | Place landmark at the posterior junction of RV and LV in a mid-ventricular slice | 15(0) subjects | 53 subjects | ED, SAX/LAX | LVVepi: 0.99, 0.94, 27.8 (CC&LRC) LVVendo: 0.99, 0.85, 6.62 (CC&LRC) | P2P: epi: 1.27–1.85, endo: 1.34–2.05 |
Kermani et al. [115] | Cine | 3D | 3D active mesh model | – | Synthetic + 6(1) real sequences | – | Multiphase, SAX | LV volume error: 3.77 ± 1.67 % LVM error: 5.26 ± 1.71 % | – |
Kurkure et al. [90] | Cine | 2D | Fuzzy connectedness based region growth | Manual segmentation on 3 ED mid-ventricle slices per subject | 20 (15) subjects | – | Multiphase, SAX + LAX | LVESV error: 8.82 ± 11.91/−3.80 ± 6.99 mL LVEDV error: −16.02 ± 19.69/−2.96 ± 7.57 mL LVEF error: 1.57 ± 5.17 %/1.83 ± 5.70 % | endo: 0.86 ± 0.12 (DC) |
Spottiswoode et al. [112] | DENSE | 2D + T | Motion trajectory guided contour propagation | Draw the initial myocardial contours on any frame | 6(0) subjects | – | Multiphase, SAX/LAX | – | Radial C2C: epi: 1.01 ± 0.23, endo: 1.29 ± 0.34 |
Suinesiaputra et al. [124] | Cine | 2D | ICA statistical model based detection and classification | Manual segmentation on training sets | 45(45) subjects | 44(0) volumes | ED + ES, SAX | – | – |
Constantinides et al. [101] | Cine | 2D | GVF based deformable model + fuzzy k-means papillary muscle detection | Place a point at the centre of LV and a point at the upper intersection of LV and RV | 15(12) subjects | 15(12) subjects | Multiphase, SAX | LEF: 0.97, 1.00, 1.60 (CC&LRC) LVM: 0.88, 0.80, 31.51 (CC&LRC) | APD: epi: 2.35 ± 0.57, endo: 2.04 ± 0.47 DC: epi: 0.92, endo: 0.89 |
Chen et al. [114] | TA | 3D + T | Deformable model based motion tracking | – | 17(11) subjects | – | Multiphase, SAX + LAX | – | – |
Cousty et al. [91] | Cine | 3D + T | Morphological region growth + watershed cuts | Specify a single point located at the centre of cavity | 18(18) subjects | – | Multiphase, SAX | – | P2S: epi:1.55 ± 0.23, endo: 1.42 ± 0.36 |
Lee et al. [84] | Cine | 2D | Region growth with iterative thresholding + active contours | Choose mid-ventricular slice | 38 patients (339 images) | – | Multiphase, SAX | LVVepi: 0.98 (CC), 2.0 ± 13.0 mL (error) LVVendo: 0.99 (CC), 2.9 ± 6.2 mL (error) LVM: 0.90, 0.93, 10.37 (CC&LRC), −0.9 ± 16.5 g (error) | – |
Schaerer et al. [107] | Cine | 2D + T | Deformable elastic template + temporal constraints | Specify a point at the centre of cavity in the ED frame | 15(15) subjects | – | Multiphase, SAX | LV volume error: −12–57 % LVEF error: 1–6 % LVM error: 4–35 % | APD: epi: 3.14 ± 0.33, endo: 2.97 ± 0.38 DC: epi: 0.92 ± 0.02, endo: 0.87 ± 0.04 |
Zhu et al. [130] | Cine | 3D + T | Propagation based subject-specific dynamic model | Manual segmentation on first frame can be required | 22(0) subjects | Leave 1 out | Multiphase, SAX | LVV error: −2.3 to 0.5 mL | MAD: epi:1.27 ± 0.18, endo: 0.69 ± 0.13 HD: epi:1.72 ± 0.15, endo: 1.47 ± 0.16 |
Cordero-Grande et al. [118] | Cine | 3D + T | Markov random field based deformable model | – | 43(43) subjects | – | Multiphase, SAX | LVESV: −7.19, 1.05 (LRC); error: −3.3 ± 7.2 mL LVEDV: −1.59, 0.99 (LRC); error: −3.6 ± 8.2 mL LVEF: −2.23, 1.07 (LRC); error: 1.5 ± 3.3 % LVM: 2.22, 1.06 (LRC); error: 8.2 ± 11.6 g | S2S: epi: 1.22 ± 0.17, endo: 1.37 ± 0.20 |
Huang et al. [86] | Cine | 2D | Thresholding + edge detection + radial region growth | Choose mid slice and manual correction can be required | 45(36) subjects | – | Multiphase, SAX | – | APD: epi: 2.22 ± 0.43, endo: 2.16 ± 0.46 DC: epi: 0.93 ± 0.02, endo: 0.89 ± 0.04 |
Lekadir et al. [126] | Cine | 3D + T | PDM + local spatial–temporal descriptor | Place 4 landmarks | 50 subjects | Cross validation | Multiphase, SAX + LAX | – | 1.46 ± 0.35 |
Brien et al. [127] | Cine | 3D + T | ASM + global contour optimisation | Manual segmentation on training sets |
n subjects | 33-n subjects | Multiphase, SAX + LAX | LVVepi: 0.97–0.99 (CC) LVVendo: 0.88–0.95 (CC) | – |
Ammar et al. [88] | Cine/LE/P | 2D | Thresholding + level-set | – | 18(18) subjects | – | Multiphase, SAX | – | – |
Ayed et al. [133] | Cine | 2D | Subject-specific model + max-flow optimisation | Manual segmentation on first frame | 20 subjects (2280 slices) | – | Multiphase, SAX | LVVendo: 0.99 (CC) Myocardial volume: 0.81 (CC) | DC: endo: 0.92 ± 0.03, myocardium: 0.82 ± 0.06 |
Khalifa et al. [116] | Cine | 3D | Level-set based geometric deformable model with prior | Manual segmentation on training sets | 26(26) subjects | 1/3 of total | Multiphase, SAX | – | APD: epi: 0.87 ± 0.52, endo: 1.21 ± 1.29 DC: epi: 0.96 ± 0.02, endo: 0.91 ± 0.07 |
Ringenberg et al. [102] | Cine + LE | 2D | Thresholding and Canny edge detection based ROI extraction + GVF snake | – | 5(5) subjects | – | Multiphase, SAX | – | P2C: Cine: epi: 1.45 ± 0.65, endo: 1.25 ± 0.39; LE: epi: 1.95 ± 0.85, endo: 1.73 ± 0.69 |
Eslami et al. [135] | Cine | 3D + T | Retrieval closet subject with guided random walks | Provide myocardial and background seeds on ED frame | 104(73) subjects | – | Multiphase, SAX | LVESV: 0.98, 0.96, 2.10 (CC&LRC) LVEDV: 0.98, 0.92, 1.58 (CC&LRC) LVEF: 0.96, 0.99, 1.61 (CC&LRC) LVM: 0.95, 1.04, 3.40 (CC&LRC) | P2C: epi: 1.48 ± 0.44, endo: 1.54 ± 0.31 DC(%): endo: 83.25 ± 3.05 (ED), 84.69 ± 4.17 (ES); epi: 80.15 ± 3.89 (ED), 79.65 ± 5.04 (ES) |
Hu et al. [93] | Cine | 2D | GMM (EM) + region restricted dynamic programming | – | 45(36) subjects | – | Multiphase, SAX | LVEF: 0.94, 1.01, 2.76 (CC&LRC) LVM: 0.82, 0.90, 9.9 (CC&LRC) | APD: epi: 2.21 ± 0.45, endo: 2.24 ± 0.40 DC: epi:0.94 ± 0.02, endo: 0.89 ± 0.03 |
Lu et al. [87] | Cine | 2D | Optimal thresholding + FFT + multiple seeds region growth | Choose mid-slice and manual correction can be required | 133(96) subjects | – | Multiphase, SAX | LVEDV: 0.98 (CC), LVESV: 0.98 (CC) LVEF: 0.90 (CC), LVM: 0.88 (CC) | APD: epi: 0.92, endo: 2.08 DC: epi: 0.94, endo: 0.90 |
Nambakahsh et al. [134] | Cine | 3D | Convex relaxed + distribution matching with priors | Specify a single point on target region (cavity or myocardium) | 20 subjects (400 volumes) | Leave 1 in | Multiphase, SAX | LVVepi: 0.91 (CC) LVVendo: 0.88 (CC) | DC: epi: 0.70 ± 0.01, endo: 0.80 ± 0,02 |
Roohi and Zoroofi [128] | Cine | 3D + T | Kernel PCA based 3D ASM | Manual segmentation on training sets | 33 subjects (7980 slices) | Leave 1 out | Multiphase, SAX | LVVepi: 0.99, 1.92 (LRC) LVVendo: 1.00, 1.61 (LRC) | – |
Wei et al. [117] | Cine + LE | 3D | Propagate contours prior from cine to LE + deformable model | Exclude the most basal and apical slices and selects one 4-chamber and one 2-chamber LAX slices from LE images | 12 patients, 4 simulated phantom data | – | One phase, SAX + LAX | – | epi: 0.67 ± 0.41, endo: 0.73 ± 0.49 DC(%): 98.05 ± 0.07 |
Woo et al. [99] | Cine | 2D | Coupled level-set + dual shape constraint | Choose centre of endocardium and its boundary by 2 clicks on mid-slice at ED | 15 subjects | – | Multiphase, SAX | LVESV: 68 ± 49 mL (Grd 69 ± 45 mL) LVEDV: 139 ± 44 mL (Grd 139 ± 41 mL) LVEF(%): 54 ± 16 (Grd 55 ± 19) | DC: 0.89 ± 0.03 |
Wu et al. [103] | Cine | 2D | GVC based parametric active contour | – | 126(0) + 45(45) images | – | Multiphase, SAX | – | MAD: epi: 5.18 pixels, endo: 5.06 pixels |
Afshin et al. [14] | Cine | 2D | Image feature + LDA + linear SVM classification | Specify initial segmentation and anatomical landmarks on the first SAX slice | 58(37) subjects | Threefold cross validation | Multiphase, SAX | Classification accuracy: 86.09 % | – |
Alba et al. [131] | Cine/LE | 3D | Intensity based graph-cuts + inter-slice and shape constraint | – | 15 cine + 20 LE patients | – | Multiphase, SAX | – | P2P: Cine: epi: 2.58 ± 0.39, endo: 2.76 ± 0.53; LE: epi: 2.38 ± 0.53, endo: 1.83 ± 0.50 DC: Cine: 0.92 ± 0.04, LE: 0.81 ± 0.05 |
Auger et al. [113] | DENSE | 3D | Displacement based contour propagation + model fitting | Specify guide points on myocardial borders on 3 SAX slices (apical, mid, and basal) | 4(0) subjects | – | Multiphase, SAX | – | DC: 0.92 |
Qin et al. [129] | Cine | 2D | Feature competition + sparse model + incremental learning | Manual segmentation on first frame | 33 subjects (mid slices) | Leave 1 out | Multiphase, SAX | – | P2C: epi: 1.44 ± 0.36, endo: 1.75 ± 0.50 DC: epi: 0.95 ± 0.01, endo: 0.90 ± 0.03 |
Queiros et al. [89] | Cine | 3D + T | B-spline explicit active surface + sequential thresholding + EM | Choose basal and apical slices by 2 clicks | 45(36) subjects | – | Multiphase, SAX | LVEDV: 0.985, 0.99, −1.04 (CC&LRC) LVESV: 0.988, 1.026, −6.903 (CC&LRC) LVM: 0.951, 1.04, 0.69 (CC&LRC) LVEF: 0.976, 1.10, −1.63 (CC&LRC) | APD: epi: 1.80 ± 0.41, endo: 1.76 ± 0.45 DC: epi: 0.94 ± 0.02, endo: 0.90 ± 0.05 |
Bai et al. [97] | Cine | 3D | Multi-atlas + augmented feature + SVM classification | Place 5 landmarks on ED frames in the target and atlas | 83 subjects | Leave 1 out | Multiphase, SAX | LVESV error: 9.3 ± 9.9 mL LVEDV error: 8.9 ± 8.2 mL LVM error: 11.9 ± 12.4 g | DC: 0.807 |
References | Mode | Dim | Fundamental principles | User interaction | Test cases | Training sets | Materials | CC and LRC | HD (mm) | DC |
---|---|---|---|---|---|---|---|---|---|---|
Maier et al. [136] | Cine | 3D + T | Region-growing (watershed) + graph-cut | Specify the midline of RV wall in ED slices or 2 points on ED basal slice for registration | 16(16) subjects | 16(16) subjects if atlas is in use | Multiphase, SAX | RVESV: 0.96, 1.06, 6.73 RVEDV: 0.99, 1.06, 1.02 RVEF: 0.86, 0.07, −0.06 | endo: 14.75 ± 0.40 (ES), 9.21 ± 0.29 (ED) | endo: 0.69 ± 0.02 (ES), 0.84 ± 0.01(ED) |
Ou et al. [142] | Cine | 3D | Atlas registration based propagation + label fusion | – | 16(16) subjects | 16(16) subjects | ED + ES, SAX | – | epi: 21.91 ± 18.92 (ES), 19.21 ± 18.50 (ED) endo: 20.44 ± 17.80 (ES), 18.77 ± 18.29 (ED) | epi: 0.60 ± 0.30 (ES), 0.69 ± 0.28 (ED) endo: 0.53 ± 0.32 (ES), 0.65 ± 0.30 (ED) |
Wang et al. [117] | Cine | 3D + T | X–Y direction spatial morphological patterns + Z and temporal refinement | – | 16(16) subjects | – | Multiphase, SAX | RVESV: 0.80, 1.56, 2.30 RVEDV: 0.87, 1.37, 5.14 RVEF: 0.29, 0.48, 0.19 | epi: 27.58 ± 24.82 (ES), 21.45 ± 25.14 (ED) endo: 27.99 ± 24.97 (ES), 22.89 ± 25.01 (ED) | epi: 0.55 ± 0.36 (ES), 0.70 ± 0.34 (ED) endo: 0.50 ± 0.34 (ES), 0.63 ± 0.32 (ED) |
Zuluaga et al. [143] | Cine | 2D | Atlas based coarse-to-fine segmentation + label fusion | – | 16(16) subjects | 16(16) subjects | ED + ES, SAX | RVESV: 0.97, –, – RVEDV: 0.96, –, – | epi: 11.81 ± 9.46 (ES), 10.23 ± 7.22 (ED) endo: 11.41 ± 10.49 (ES), 9.77 ± 7.88 (ED) | epi: 0.77 ± 0.23 (ES), 0.86 ± 0.13 (ED) endo: 0.72 ± 0.27 (ES), 0.83 ± 0.17 (ED) |
Bai et al. [144] | Cine | 3D | Multi-atlas registration + label fusion | Specify a few landmarks on ED slices for registration | 16(16) subjects | 16(16) subjects | ED + ES, SAX | RVESV: 0.98, 0.67, 12.13 RVEDV: 0.99, 0.87, 17.86 RVEF: 0.92, 0.57, 0.29 RV mass: 0.91,0.82, 1.35 | epi: 11.72 ± 5.44 (ES), 7.93 ± 3.72 (ED) endo: 11.16 ± 5.53 (ES), 7.70 ± 3.74 (ED) | epi: 0.77 ± 0.17 (ES), 0.88 ± 0.08 (ED) endo: 0.69 ± 0.25 (ES), 0.86 ± 0.11 (ED) |
Nambakhsh et al. [141] | Cine | 3D | Prior distribution matching + convex relaxation | Specify the centroid of LV and a small closed region inside RV cavity in the middle slice | 32(32) subjects | Leave 1 in | ED + ES, SAX | RVESV: 0.79, 1.05, 52.04 RVEDV: 0.81, 1.02, 36.48 RVEF: 0.28, 0.38, 0.10 | endo: 23.19 ± 9.71 (ES), 17.76 ± 7.73 (ED) | endo: 0.48 ± 0.25 (ES), 0.67 ± 0.19 (ED) |
Grosgeorge et al. [145] | Cine | 2D | Distance map-based SSM + registration + graph cut | Place 2 anatomical landmarks on the ventricular septum | 16(16) subjects | 16(16) subjects | ED + ES, SAX | – | – | endo: 0.70 ± 0.22 (ES), 0.83 ± 0.15 (ED) |
Mahapatra [140] | Cine | 2D/3D | Super-pixel or super-voxel classification by random forest | – | 32 datasets | Leave 1 out | Multiphase, SAX | – | endo: 6.7 | endo: 0.93 |
Oghli et al. [146] | Cine | 2D | Robust PCA shape based deformable model | Manual segmentation on training sets | 30(30) slices | 30 binary shapes | ED + ES, SAX | – | – | – |
Ringenberg et al. [138] | Cine | 2D | PCA window constraints + accumulator thresholding | Manual segmentation on training sets | 32(32) subjects | 16(16) subjects | ED + ES, SAX | RVESV: 0.95, 1.02, 10.16 RVEDV: 0.98, 1.10, −6.64 RVEF: 0.78, 0.83, 0.02 RV mass: 0.97, 1.10, −2.77 | epi: 11.52 ± 7.70 (ES), 8.02 ± 5.96 (ED) endo: 10.71 ± 7.69 (ES), 7.69 ± 6.03 (ED) | epi: 0.82 ± 0.13 (ES), 0.90 ± 0.08 (ED) endo: 0.77 ± 0.18 (ES), 0.88 ± 0.11 (ED) |
Punithakumar et al. [139] | Cine | 2D + T | Moving mesh propagation by point-to-point correspondence | Manual segmentation on a single initial frame | 48(48) +23(23) subjects | – | Multiphase, SAX | – | epi: 8.08 ± 3.80 endo: 7.72 ± 3.97 | epi: 0.87 ± 0.08 endo: 0.83 ± 0.13 |
References | Mode | Dim | Fundamental principles | User interaction | Test cases | Training sets | Materials | CC and LRC | LVendo/LVepi/RV distance (mm) |
---|---|---|---|---|---|---|---|---|---|
Mitchell et al. [158] | Cine | 2D | ASM + AAM | Manual segmentation on training sets | 60(27) mid-ventricle slices | 102(33) mid-ventricle slices | ED, SAX | LVepi: 0.96, 0.90, 0.41 LVendo: 0.96, 1.04, −0.55 RV: 0.90, 0.97, −0.22 | P2P(signed): 0.22 ± 1.90/−0.01 ± 1.92/−0.32 ± 2.80 |
Ordas et al. [157] | Cine | 2D | ASM + invariant optimal features | Manual segmentation on training sets | 74(61) subjects | 21(13) subjects | Multiphase, SAX | – | P2C: 1.80 ± 1.74/1.52 ± 2.01/1.20 ± 1.74 |
Sermesant et al. [147] | Cine | 3D + T | Deformable biomechanical mesh registration + tracking | Choose reasonable mesh size | 2 sequences | – | Multiphase, SAX + LAX | – | – |
Lorenzo-Valdes et al. [155] | Cine | 3D + T | 4D probabilistic atlas + MRF + EM algorithm | Manual segmentation on training sets | 14(0) + 10(10) subjects | Leave 1 out | Multiphase, SAX | LVVepi: 0.92, 1.18, 7.0 LVVendo: 0.96, 0.92, −3.42 RV volume: 0.92, 0.90, 15 | P2C: 2.21 ± 2.22/2.99 ± 2.65/2.89 ± 2.56 |
Rougon et al. [148] | Cine + TA | 2D | Dense motion estimation + non-rigid propagation from ED | – | 12 subjects | – | Multiphase, SAX + LAX | – | – |
Hautvast et al. [149] | Cine | 2D | Automatic contour propagation from ED slices to ES slices | Segment an ED frame as initialisation | 69(69) SAX slices + 38(38) LAX slices | – | Multiphase, SAX/LAX | LVESV: SAX: 0.98, 1.03, −2.08; LAX: 0.93, 0.92,13.72 LVSV: SAX:0.71, 0.78, 9.74; LAX: 0.68, 0.61, 11.19 LVEF: SAX: 0.78, 0.99, 2.08; LAX: 0.76, 0.59, 12.43 | SAX (ES): 2.23 ± 1.10/1.84 ± 1.04/2.02 ± 1.21 LAX (ES): 1.82 ± 0.61/0.92 ± 0.42/– |
Cocosco et al. [150] | Cine | 3D + T | Binary voxel classification + thresholding + region-growing | Choose basal slices | 32(32) subjects | – | Multiphase, SAX | LVVendo: 0.97, 0.94, 15.7 RV volume: 0.97, 1.11, 17.9 LVEF: 0.94, 0.98, −7.58 RVEF: 0.71, 0.60, 14.3 | – |
Zhang et al. [159] | Cine | 3D + T | ASM + AAM | Fitting the mean shape prior to the first frame as initialisation | 25(0) + 25(25) subjects | Fivefold cross-validation | Multiphase, SAX + LAX | LVVendo: 0.98, 0.97, 7.1 RV volume: 0.96, 0.95, 3.0 LVM: 0.76, 0.88, 13.9 | P2S (normal): 1.67 ± 0.30/1.81 ± 0.40/2.13 ± 0.39 P2S (TOF): 1.71 ± 0.45/1.97 ± 0.58/2.92 ± 0.73 |
Grosgeorge et al. [151] | Cine | 2D | Region-based level-set | – | 59(59) subjects | – | ED + ES, SAX | – | P2C (ED):2.33–3.52/–/2.27–3.28 P2C (ES): 2.27–5.00/–/1.85–3.50 DC (ED): 0.67–0.82/–/0.46–0.80 DC (ES): 0.46–0.70/–/0.25–0.59 |
Mahapatra [153] | Cine | 2D | Single shape prior + graph-cut | Identify myocardium, LV and RV in the first frame | 30(30) subjects | – | Multiphase, SAX | – | HD: 1.8 ± 0.4/1.9 ± 0.3/2.0 ± 0.3 DC (%): 91.7 ± 1.1/91.6 ± 0.9/92.2 ± 1.2 |
Wang et al. [154] | Cine | 3D | Context-specific reinforcement learning | Place points on the correct contour during segmentation | 60(0) + 21(21) subjects | 15 subjects when segmenting RV | Multiphase, SAX | – | C2C (ED): 0.91 ± 0.18 (healthy LVendo)/1.73 ± 0.64 (healthy RV)/1.15 ± 0.25 (HCM LV) C2C (ES): 1.01 ± 0.20 (healthy LVendo)/2.32 ± 0.96 (healthy RV)/1.17 ± 0.24 (HCM LV) |
Bai et al. [156] | Cine | 3D | Multi-atlas registration + patch based probabilistic label fusion | Manual segmentation on training sets | 28(0) subjects | Leave 1 out | Multiphase, SAX | – | Average: 1.26/1.49/1.68 Maximum: 7.27/9.35/12.23 DC: 0.915/0.824/0.886 |
Wang et al. [15] | Cine | 2D | Prior probability model + direct area estimation | Place 2 landmarks on each slice in first frame | 56 subjects (3360 slices) | Leave 1 out | Multiphase, SAX | LVVendo: 0.985, –, – RV volume: 0.957, –, – LVEF: 0.966, –, – RVEF: 0.807, –, – | – |
Zhen et al. [16] | Cine | 2D | Direct estimation by multiscale deep networks and regression forest prediction | Manual segmentation on training sets | 100 subjects (6000 slices) | Unsupervised feature learning: 47 subjects (2820 slices) Supervised learning: Leave 1 out validation in 100 subjects | Multiphase, SAX | LVVendo: 0.921, –, – RV volume: 0.908, –, – | – |
Alba et al. [160] | Cine | 3D | PDM based feature searching + model fitting in various pathologies | Specify a few landmarks | 20 normal as reference + 40(40) subjects | Leave 1 out | ED, SAX | – | P2S: pulmonary hypertension: 2.60 ± 0.34; hypertrophic cardiomyopathy: 2.57 ± 0.46 |
References | Protocol | Dimension | Fundamental principles | User interaction | Test cases | Training sets | Accuracy |
---|---|---|---|---|---|---|---|
John and Rahn [161] | F | 2D | Thresholding + subdivision (narrow cuts) + region merging | Final segmentation positive and negative marking | 20 subjects | – | – |
Karim et al. [163] | LE | 2D | 3D probabilistic atlas construction + MRF based energy function minimisation within Voronoi framework | Choose 3 or 4 landmarks on each training image | 10(10) volumes | 20(20) volumes | Mean slice overlap: 0.90 |
Kutra et al. [164] | Cine/LE | 3D | Multi-model based fitting + SVM based optimal model selection | Manual segmentation on training sets | 59(47 %) subjects | Leave 1 out validation | P2S: Normal: 0.87 mm; CLT: 0.81 mm; RMPV: 0.79 mm |
Zhu et al. [162] | LE | 2D | Local seed region searching + region growth with prior | Manual segmentation on training sets | 64(64) volumes | 16 volumes | DC: 0.79 ± 0.05, Volume overlap: 0.65 ± 0.07, HD: 14.40 ± 3.65 mm, S2S: 2.79 ± 2.84 mm |
References | Mode | Dim | Fundamental principles | User interaction | Test cases | Training sets | Materials | S2S (mm) | ||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
LVepi
| LVendo
| RVepi
| RVendo
| LA | RA | Whole mesh | ||||||||
Makowski et al. [165] | Cine | 2D | 2-phase active contour (Balloon + Snake) | Place the initial contour | 70 slices | – | Multiphase, SAX + LAX | – | ||||||
Lotjonen et al. [166] | Cine | 3D | SSM + non-rigid registration | 3D surface fitting to create the prior shape model | 25(0) subjects | Leave 1 out | Multiphase, SAX + LAX |
2.77 ± 0.49
| 2.01 ± 0.31 |
2.77 ± 0.49
| 2.37 ± 0.50 | 2.56 ± 0.88 | 2.93 ± 1.30 | 2.53 ± 0.70 |
Koikkalainen et al. [167] | – | 3D | Artificial training sets enlargement for SSM | 3D surface fitting to create the prior shape model | 25(0) subjects | Leave 1 out | ED, SAX + LAX |
1.87 ± 0.63
| 1.46 ± 0.30 |
1.87 ± 0.63
| 2.26 ± 0.46 | 2.28 ± 0.63 | 3.22 ± 1.62 | 2.06 ± 0.55 |
Wierzbicki et al. [168] | Cine | 3D + T | PCA based template registration + motion extraction | Manual segmentation on training sets | 10(0) subjects | Leave 1 out | Multiphase, SAX + LAX | 3.4 ± 0.9 |
3.5 ± 1.1
| – | 3.2 ± 0.7 | 3.5 ± 1.1 | 4.2 ± 1.5 | |
Peters et al. [169] | Cine | 3D | Mesh registration + simulated search for boundary detection | 3D surface fitting to create the prior shape model | 42(42) volumes | Fourfold cross-validation | ED, SAX + LAX | 0.83 ± 1.17 | 0.69 ± 1.13 | – | 0.74 ± 0.96 | 0.72 ± 1.14 | 0.63 ± 0.95 | 0.76 ± 1.08 |
Zhuang et al. [170] | Cine | 3D | Multi-atlas propagation + refinement + label fusion | Manual segmentation on training sets | 37(19) volumes | 10 reference shapes | ED, SAX + LAX | 2.32 ± 0.82 | 1.47 ± 0.32 | – | 2.13 ± 0.70 | 2.38 ± 1.14 | 2.22 ± 0.75 | 2.14 ± 0.63 |
Zuluaga et al. [171] | Cine | 3D | Multi-atlas propagation + refinement + label fusion | Manual segmentation on training sets | 22 subjects | Leave 1 out | ED + ES, SAX | DC: LV volume: 0.95; RV volume: 0.92; LA volume: 0.92; RA volume: 0.89; Myocardium: 0.87; Aorta: 0.86 | ||||||
Zhen et al. [17] | Cine | 2D | Multi-output regression with random forest | Manual segmentation on training sets | 125 subjects | Leave 1 out | Multiphase, SAX | CC: LV volume: 0.91; LA volume: 0.87; RV volume: 0.88; RA volume: 0.86 |
The anatomy of the heart
MRI protocols
Indices of cardiac function
LV quantification
RV quantification
LA quantification
RA quantification
Cardiac segmentation
Segmentation methodologies
LV
Thresholding and region-growing
Pixel or voxel classification
Active contours
Strong prior based techniques
Direct estimation
RV
Image-driven techniques
Model-driven techniques
Bi-ventricle
Image-driven techniques
Model-driven techniques
Direct estimation
LA
Image-driven techniques
Model-driven techniques
Whole heart
Image-driven techniques
Model-driven techniques
Direct estimation
Discussion
Choice of segmentation techniques
Segmentation of the whole heart
Segmentation of large-scale CMR datasets
Segmentation of abnormal cases
Clinical translation
Name | Producer | Use | Website |
---|---|---|---|
CAAS MRV | Pie Medical Imaging | C | piemedicalimaging.com |
CAIPI | Mevis Fraunhofer | R | mevis.fraunhofer.de |
Corridor4DM | INVIA (Siemens) | C | inviasolutions.com |
CMRtools | Cardiovascular Imaging Solutions | C/R | cmrtools.com |
CVI42 | Circle Cardiovascular Imaging | C | circlecvi.com |
GIMIAS Cardio Suite | CISTIB | R | gimias.org |
Heart IT | Heart Imaging Technologies | C | heartit.com |
iNtuition Cardiac | TeraRecon | C | terarecon.com |
PiA CMR | Precision Image Analysis | C | piamedical.com |
Qmass | Medis | C | medis.nl |
Segment CMR | Medviso | C | medviso.com |
Ziostation MR Cardiac Function | Qi Imaging | C | qiimaging.com |