In order to identify the health status of the heart using images of LV following evaluation parameters were obtained after the segmentation process.
In order to identify the health status of the heart using images of LV following evaluation parameters were obtained after the segmentation process. - Intersection of Union (IoU), DM, HD, Jaccard Score, Conformity Index CI, RPC, CV, Percentage of good contours, Actual average perpendicular distance, cardiac parameters, R-squared Score. All three algorithms, CNN based U-Net Model, VGG 16, and ResNet 152 are analyzed for the above parameters.
Following are the expected values for the evaluation matrices -
- IOU and DM > 0.70.
- CI and JS score > 0.7
- PGC, AVPD, EDV, ESV and EF >=70% and <= 90%
- R-squared >= 0.7 and <=0.9.
Scores of the above-mentioned factors have been enlisted in the tables and figures given below. We can compare the intersection of union value of all the algorithms in the given image.
From Figure 1, we can see that model based on u-net architecture is performing better than VGG-16 and ResNet152 algorithms. Figure 2 is expressing the DM scores using the 3 algorithms. It is also observed that U-Net model obtains the highest scores for DM. The next Figure 3 is expressing the R-squared values obtained from the implementation of 3 DL models for segmentation of images and the results state that U-Net outperforms other two models with a marginal difference.
Figure 4 is expressing the results of HD scores for three proposed schemes namely RESNET, VGG 16 and U-Net model for image segmentation. The HD scores should be minimal as per the statistical significance and U-Net model outperforms others with minimal HD scores. Figure 5 expresses the CI values which should be in the range of 0.7 to 0.99 and maximum value is considered to be the best. It is observed that U-Net again gives the best scored for CI, followed by RESNET and VGG 16.
Figure6 is also expressing the comparison of EDV values obtained by the three algorithms from segmentation of images and it can be inferred that all three techniques work well in interpretation of results and U-Net performs the best. Figure 7 expresses values for EF and all 3 algorithms are able to depict EF values from the segmented images of heart.
The evaluation matrix has been calculated for each of the algorithms. The same has been tabulated as the following.
Table.1 Comparison of various Parameters w.r.t. each algorithm
Parameters/ Algorithms
|
RESNET
|
VGG 16
|
CNN-based U-Net Model
|
Coefficient of variation CV
|
3.3211
|
1.4274
|
1.2566
|
POGC
|
68.739
|
74.9547
|
93.1822
|
Actual APD
|
69.9788
|
68.9992
|
88.5394
|
Predicted APD
|
66.2137
|
73.8518
|
94.0271
|
It is clear from all the results revealed above that U-Net outperforms RESNET and VGG 16. It is clear that IOU and DM parameters show values slightly greater than 0.70. CI and JS scores are greater than 0.7 PGC, AVPD, EDV, ESV and EF do not show the expected range (>=70% ). R-squared >= 0.7 and <=0.9. This shows that the VGG-16 model does not show ESV and EF with accuracy. IOU shows the value of less than 0.70. CI and JS scores are greater than 0.7.EDV, ESV are less than 70%. R-squared >= 0.7 and <=0.9. This shows that the ResNet model shows less accuracy as compared to the CNN-based U-Net model in segmentation. Let us check the comparison of the coefficient of variation of the all three algorithms. Figure 10 displays the bar-chart for comparing the values of coefficient of variation of all the three algorithms.
Lesser the coefficient of variation better is the performance of the model. In our analysis model with u-net architecture has lesser coefficient of variation or has the coefficient of variation value in proper range i.e. 0-1 as shown in Figure 8. But for other algorithms like VGG-16 and ResNet152, we can see that the coefficient of variation value is incorrect. The correlation plot and Bland Altman plot (BAP) have been plotted for the EF, EDS, and ESV for three deployed algorithms. BAPs are primarily used to evaluate the contract between two measurement techniques or two different instruments. In order to identify any systematic difference between the measurements or possible outliers, BAP is used. The following diagrams show the correlation plot and the BAP for factors EF, EDS, and ESV for all the three algorithms. We have depicted the Correlation and BAPs for cardiac-parameters for CNN based U-Net Model shown in Figure 9.
The correlation plot and Bland Altman plot (BAP) have been plotted for the EF, EDS, and ESV for the CNN-based U-Net Model. BAPs are used to evaluate the contract between two measurement techniques or two different instruments. Correlation and BAPs for cardiac-parameters for VGG 16 are as shown in the Figure 9
The correlation plot and Bland Altman plot (BAP) have been plotted for the EF, EDS, and ESV for the VGG 16 Model. BAPs are used to evaluate the contract between two measurement techniques or two different instruments. Correlation and BAPs for cardiac parameters for ResNet-152 are shown in Figure 10
The correlation plot and Bland Altman plot (BAP) have been plotted for the EF, EDS, and ESV for the ResNet-152 Model. BAPs are used to evaluate the contract between two measurement techniques or two different instruments. The results of segmentation for all three algorithms have been depicted in the following figures. The following figure (Figure 10) shows results obtained after segmentation using CNN based U-Net Model
The following figure (Figure 12) shows results obtained after segmentation using VGG 16.
The following figure (Figure13) shows the results obtained after segmentation using ResNet 152.
The results depict that CNN U-Net and VGG16 show better accuracy while performing the LVS. ResNet-152 does not perform up to the expected level of accuracy in the results. The segmentation is carried out using CNN U-Net and VGG16 and ResNet-152. It involves complexities as the LV is difficult to identify in the image because of the position and blood flow. After segmentation, CNN U-Net and VGG16 are able to identify the correct position of LV and LV is clearly separable from the rest of the MRI image.
3.4 Discussion
Following are the details regarding the evaluation metrics -
The r-squared score, dice metric, and IoU are the major factors obtained from the matrices. Their values are between 0 to 1. The proposed model is generating an accurate value of R-squared factor. It is close to 1. IoU and Dice metric values should be close to 1. When the three algorithms were compared, CNN based U-Net Model is found to be performing better than the other two algorithms. We can conclude the CNN-based based U-Net Model shows a percentage of good contours up to 89%, thus identifying the suffered area in the heart. The IoU metric is also referred to as the Jaccard Score, quantifies the percent overlap between the actual area and the prediction area in the image.
Bland-Altman plots reveal the following outcomes-
- CNN based U-Net Model shows a positive relationship between the actual values and the predicted values. This signifies that the predicted values are closer to the actual ones.
- VGG 16 and ResNet 152 do not show positive relation. Hence are not predicting the values to the expected accuracy.
When the above three models are applied to the problem statement, CNN based U-Net Model produces the most accurate results.
Chronic CVD and ailments need frequent hospital visits and constant observation. The Blockchain is a secure, distributed database technology that does not need any third-party invasion which makes it safer to store cardiac images. The images and the records can be used to remotely manage chronic issues, including self-care before hospital visits as well as post-hospital visit care. Blockchain databases can securely store cardiac images and the above-mentioned algorithms can be used to accurately segment them in order to identify the important cardiac parameters. These parameters can be used to diagnose cardiac diseases.