In comparison with existing published review papers [15] presented classified segmentation methods based on strong, weak, and prior knowledge, [16] focused on deep learning techniques using cardiac image segmentation, including modalities such as CMR, CT, ultrasound, [17] is the study of spatio-temporal image information and evaluation of clini- cal usability using deep learning applications. Our review focuses on different parameters and the limitations asso- ciated with manual, semi-automatic, and fully-automatic approaches used for CMR-LV contour segmentation and quantitative measurements of cardiac function.
3.1 Manual Segmentation Approach
Figure 4 shows a manual segmentation approach, first in- troduced by a 17-segments model developed by American Heart Association (AHA) [18]. American Heart Association (AHA) presented an anatomical structure and recommended a division into 17-segments of the right atrium,pulmonary atrium, RV, left atrium, LV, coronary arteries, aorta, and cardiac veins. This model is further divided into three ma- jor domains, namely basal, cavity, and apex in a series of numbers. From digits 1 to 6, index the basal information; 7 to 12 are the cavity information; and 13 to 17 are the apex information. Basically, a manual approach requires an expert for LV contour tracing. There are some studies that used manual segmentation approaches for tracing the contour Including, [19] presented hand-operated of LV and Right Ventricle (RV) contour using quantitative analyses of global and regional information of segmented contours [20]. Further, the study presented a hand-operated tracing method that visualised the anatomical view of annotated neighbour regions of detected LV contours using reference ranges to permit differentiation between normal and remodel features. In clinical practice, this method is famous because of its advantages for assessing cardiac function. Figure 4 shows a manual segmentation approach, first introduced by a 17- segments model developed by American Heart Association (AHA) [18]. American Heart Association (AHA) presented an anatomical structure and recommended a division into 17-segments of the right atrium,pulmonary atrium, RV, left atrium, LV, coronary arteries, aorta, and cardiac veins. This model is further divided into three major domains, namely basal, cavity, and apex in a series of numbers. From digits 1 to 6, index the basal information; 7 to 12 are the cavity information; and 13 to 17 are the apex information. Basi- cally, a manual approach requires an expert for LV contour tracing. There are some studies that used manual segmen- tation approaches for tracing the contour Including, [19] presented hand-operated of LV and Right Ventricle (RV) contour using quantitative analyses of global and regional information of segmented contours [20]. Further, the study presented a hand-operated tracing method that visualised the anatomical view of annotated neighbour regions of detected LV contours using reference ranges to permit differentiation between normal and remodel features. In clinical practice, this method is famous because of its advantages for assessing cardiac function.
3.2 Semi-Automatic Approach
This approach promotes a systematic solution to save the experts time and produces competitively segmented contour points as compared with manually annotated contour points. Different studies were conducted in the domain of a semi- automatic approach, including atlas-based, guide point, ac- tive contours, and classifier. This approach reviewed re- lated studies made based on semi-automatic segmented LV contours using short-axis images such as modified active contours model that extracted good contours for predicting cardiac function [21]; the edge-based method for enhancing segmented LV contour information [22];; the guide point approach [23]; the radial snake approach [24]; atlas-based segmentation approach detected geometrical shape features that lead when cardiac remodeling in a LV [25, 26] and [27]; shape and motion reconstruction approach along with a spatio-temporal statistical atlas for capturing motion during the cardiac remodeling and as a result produces quantitative measurements of cardiac function after MI [28]; region- based model and active contour method used for segmenta- tion of increases size LV Endocardial contour calculated in percentage [29]. Some other methods were presented with automatic correction for quantitative solution of myocardial oedema but because of manual correction, these methods were considered semi-automatic instead. Quantifying seg- mented oedema is an open challenge for both cardiac re- searchers and computer scientists. Due to the limitations of these methods, additional research contributions are re- quired to quantify the amount of oedema. A validation study of automatic features analysis for the quantification of dead cells after MI was reported in [30], which used a combined threshold infarct sizing algorithm and a 2- standard division (2 SD) statistical method. [31] presents an oedema quantification solution using a 2-SD statistical method. In addition [32] used 2-SD to quantify ischemia disease using an automatic classifier, which identified the ischemic myocardial salvage and area at risk (AAR).
3.3 Fully Automatic Approach
The approach is divided into two categories: LV local- ization and contour segmentation, including shape feature detection [33, 34], LV segmentation and function estimation using deep learning [35, 36, 37, 38, 39]; automatic iden- tification of blood vessel using a cascading classifier [40], diffusion-based unsupervised clustering technique for My- ocardial motion patterns classification [41]; CNN and U- Net approach [42], Multi-input fusion network [43], car- diac motion measurement by used algorithm namely surface structure feature matching [44], deep earning method used deformable, level set and threshold method for automatic LV contour segmentation [45, 46, 47, 48]. Another major challenge is a region of interest (ROI) for automatic contour segmentation. In this issue, the major goal is to understand the structure of LV contour internal and external features. To overcome these issues [49] developed a hybrid-based solution and adopted two basic approaches, time-based and shape-based, which automatically localize LV in each slice with a range of 260 slices for a single patient. This method began by utilizing the shape-features of localized LV, and drawing ROI automatically to crop the localized LV ac- cording to the area of interest. Furthermore, this method is compatible with the automatic segmentation measuring function for dividing inner and outer contours cite16. Car- diac function, namely EVD, ESV, and EF assessment by automatic solution is another challenge for cardiac experts and computer scientists. The challenge includes quantitative assessment based on segmented contours. For this purpose, there have been some related studies conducted, respectfully [110] introduce a hybrid segmentation approach based on adopted methods including Gaussian-mixture and region re- striction for LV contour segmentation and assessment of car- diac function, graph cut using neighbour region of LV [111], and the graph cut approach incorporates [112, 113], random walks approach using active appearance model [114], and dynamic shape constrained random walks approach [115].
Table 2 summaries different studies conducted results against automatically segmenting LV contours using SAX images included Local Binary Fitting (LBF) Model and Dynamic Programming (DP) [50], Blind Mask Detection and Morphological Operation [51], fully Convolutional Neural Network and dynamic programming approach [52], shape prior constrained two-layer level set [53], deep learning and curriculum learning strategy [54], level set method and edge-based method [55], object detection and classification [56], Deep Convolutional Neural Network (CNN) along Level set [57] and others. Different studies were con- ducted based on good detection of contours (Endocardium (Endo) and Epicardium (Epi)) calculated in percentage. Study [49] shows a very high accuracy of Good Detected Contours (Endo = 100%, Epi = 100%) as compare with the next study [50] shows very promises results. Further in Table 2 show a comparison between Good Detected Contours with Clinical evaluation where the Dice similarity and APD calculate the average error rate among ground truth and detected contours. The major studies results were conducted based on automatic solution. The study strength represents method˜techniques used for segmenting cardiac contour. SAX represent image plan.
Table 2
Comparison of Good Detected Contours with clinical Evaluation DBN: Deep Belief Network; LBF: Local Binary Fitting; DP: Dynamic Programming; ACM: Active Contour Model; CNN: Conventional Neural Network, LSM: Level Set Method,DL: Deep Learning, DLN:Deep Learning Network, and ACM: Active contour Model, CE: Clinical Evaluation.
Source
|
Study/Strength
|
LV
|
Good Contour (%)
|
SAX
|
Error Finding
|
CE
|
Year
|
Method
|
|
|
|
|
Sunnybrook [50]
|
2014
|
LBF,DP
|
√
|
91.07
|
96.44
|
√
|
0.91
|
2.00
|
√
|
Sunnybrook [9]
|
2014
|
ACM
|
√
|
92.7
|
95.4
|
√
|
0.90
|
1.76
|
√
|
Sunnybrook [62]
|
2016
|
CNN
|
√
|
79.02
|
95.62
|
√
|
0.95
|
1.81
|
NA
|
Sunnybrook [92]
|
2016
|
LSM
|
√
|
99.14
|
99.44
|
√
|
NA
|
NA
|
√
|
Sunnybrook [51]
|
2018
|
Morphological
|
√
|
100
|
100
|
√
|
93.30
|
2.45
|
√
|
York [79]
|
2018
|
CNN
|
√
|
NA
|
NA
|
√
|
87.24
|
NA
|
NA
|
Sunnybrook [52]
|
2019
|
DP
|
√
|
96.80
|
98.4
|
√
|
0.93
|
5.00
|
√
|
ACDC [27]
|
2020
|
DLP
|
√
|
NA
|
NA
|
√
|
95.44
|
1.82
|
NA
|
ACDC[38]
|
2020
|
DL
|
√
|
NA
|
NA
|
√
|
NA
|
NA
|
√
|
Sunnybrook [53]
|
2021
|
LSM
|
√
|
100
|
100
|
√
|
0.02
|
0.01
|
√
|
HUST [54]
|
2021
|
DL
|
√
|
85.57
|
94.07
|
√
|
NA
|
NA
|
NA
|
ACDC [55]
|
2021
|
ACM
|
√
|
92
|
NA
|
√
|
0.91
|
1.5
|
√
|
The rest of the Table 3 is based on existing segmentation models including region-based, image-based, deformable- based, edge-based, active shape-based, active contour-based, Level set-based and Variational level set-based Model. The NA mean that the result is not available. Error differentiate the highest and lowest values of manual segmented ground- truth and experimental segmented contour of existing studies. Respective studies concluded clinical evaluation based on detected good contour values. Further, this Table 3 shows the accuracy in term of error between ground-truth and generated results of different studies. Good Contour shows the overall accuracy of segmented contour using basic segmentation models. The above Table 2 as compare with Table 3 show a difference in term of, where the Good Con- tour results shown Table 2 Endo and Epi in percentage(%) along-with Dice (%) and APD (mm), while in Table 3 shown overall error rate based on accuracy result of segmented contour.
Table 3
Comparison of Good Detected Contours and Clinical Assessment Based on Art of Segmentation Model S: Semi-Automatic Segmentation; A: Automatic Segmentation; Seg. Model: represents the art of the segmentation models; Good Contour: the average of detected contour value (Endocardium); Error: Dice Similarity between ground-truth and resulted contours; IBM: Image Based Model; DBM: Deformable Based Model; RBM: Region Based Model; ASM: Active Shape Based Model; ACM: Active contour Based Model; LSM: Level Set Based Model; VLSM: Variational LSM; CE: Clinical Evaluation.
Source
|
Subject
|
Seg. Model
|
Seg. Model
|
GC
|
Image Mod.
|
Error
|
CE
|
S
|
A
|
|
LV
|
SAX
|
MICCAI [21]
|
45
|
✓
|
×
|
LSM
|
91.66
|
✓
|
✓
|
NA
|
✓
|
MICCAI [50]
|
45
|
×
|
✓
|
RBM
|
93.5
|
✓
|
✓
|
0.91
|
✓
|
Sunnybrook [14]
|
45
|
×
|
✓
|
VLSM
|
94
|
✓
|
✓
|
4.84
|
✓
|
MICCAI [19]
|
45
|
✓
|
×
|
ACM
|
NA
|
✓
|
✓
|
0.97
|
✓
|
MICCAI [46]
|
45
|
×
|
✓
|
LSM
|
NA
|
✓
|
✓
|
0.93
|
✓
|
Sunnybrook [49]
|
45
|
×
|
✓
|
VLSM
|
94.50
|
✓
|
✓
|
9.19
|
✓
|
MICCAI [51]
|
45
|
×
|
✓
|
VLSM
|
93.30
|
✓
|
✓
|
NA
|
✓
|
Sunnybrook [62]
|
45
|
×
|
✓
|
DBM
|
96.00
|
✓
|
✓
|
2.45
|
✓
|
Hubei [52]
|
900
|
×
|
✓
|
EMB
|
96.08
|
✓
|
✓
|
0.90
|
✓
|
Sunnybrook [57]
|
45
|
×
|
✓
|
LSM
|
94.6
|
✓
|
✓
|
0.94
|
✓
|
MICCAI [53]
|
45
|
×
|
✓
|
LSM
|
100
|
✓
|
✓
|
0.92
|
✓
|
MICCAI [22]
|
45
|
×
|
✓
|
LSM
|
98.45
|
✓
|
✓
|
0.90
|
✓
|
CSF [27]
|
17
|
×
|
✓
|
DBM
|
98.96
|
✓
|
✓
|
0.05
|
✓
|
Heart [29]
|
18
|
✓
|
×
|
ACM
|
91.02
|
✓
|
✓
|
NA
|
✓
|
ACDC [55]
|
100
|
×
|
✓
|
ACM
|
98.14
|
✓
|
✓
|
0.90
|
✓
|
Sunnybrook [66]
|
45
|
×
|
✓
|
RBM
|
93.30
|
✓
|
✓
|
2.45
|
✓
|
Public [73]
|
660
|
✓
|
×
|
DBM
|
88.08
|
✓
|
✓
|
NA
|
✓
|
MICCAI [78]
|
45
|
✓
|
×
|
ACM
|
98.06
|
✓
|
✓
|
0.91
|
✓
|
York [79]
|
33
|
×
|
✓
|
RBM
|
87.24
|
✓
|
✓
|
NA
|
NA
|
Public [84]
|
18
|
×
|
✓
|
LSM
|
97.00
|
✓
|
✓
|
NA
|
NA
|
MICCAI [93]
|
45
|
×
|
✓
|
LSM
|
97.19
|
✓
|
✓
|
0.90
|
NA
|
MICCAI [100]
|
45
|
×
|
✓
|
RBM
|
90.06
|
✓
|
✓
|
0.88
|
NA
|
Sunnybrook [110]
|
45
|
×
|
✓
|
DBM
|
91.14
|
✓
|
✓
|
0.92
|
✓
|
Sunnybrook [122]
|
45
|
×
|
✓
|
VLSM
|
95.00
|
✓
|
✓
|
14.74
|
✓
|
Fig 5 shows a visual resulted images taken from [49], where sample shows I) detected good contours, II) edge detected technique, III) mask of good contour, IV) Distance contour mapping measurement function, V), and VI) segmented contour. The detail discussion are presented in acknowledge study [49], where weak edges information, myocardium muscle detections and contours detection methods are presented using Variational Level Set, edges detection and statistical simulation.