This chapter concludes the present work by collectively comparing the performance of the different computer-aided classification (CAC) systems designed. It also discusses the future work that involves other convolution neural network (CNN) models, such as VGGNet and ResNet-50, and other variants, such as DarkNet and NasNet. These CNN models can be used to design similar CAC systems for the binary classification of chest radiographs as designed in the present work. These CAC systems could be designed for multiclass classification such as the three-class classification of chest radiographs into Normal, Pneumonia, and COVID-19 classes.
Multiclass classification; Hybrid CAC; Handcrafted features; Feature fusion; Decision fusion; COVID-19; VGGNet; ResNet-50; DarkNet; NasNet
The present work aims at improving the diagnostic capabilities of radiologists by designing convolution neural network (CNN) model-based computer-aided classification (CAC) systems for the binary classification of chest radiographs into Normal and Pneumonia class. For this, multiple experiments have been performed for: (a) designing end-to-end pretrained CNN-based CAC systems for chest radiographs using AlexNet, ResNet18, and GoogLeNet to identify the best performing CNN model; (b) designing a hybrid CAC system for chest radiographs using the CNN model with best performance, deep feature extraction, and adaptive neuro-fuzzy classifier with linguistic hedges (ANFC-LH); (c) designing a hybrid CAC system for chest radiographs using the CNN model with best performance, deep feature extraction, and principal component analysis and support vector machine (PCA-SVM) classifier; (d) designing lightweight end-to-end pretrained CNN-based CAC systems for chest radiographs using SqueezeNet, ShuffleNet, and MobileNetV2 to identify the best performing lightweight CNN model; (e) designing a hybrid CAC system for chest radiographs using lightweight CNN model with best performance, deep feature extraction, and ANFC-LH classifier, (f) designing a hybrid CAC system for chest radiographs using lightweight CNN model with best performance, deep feature extraction, and PCA-SVM classifier. The workflow of these experiments are discussed briefly in Chapter 3.
The conclusions derived from the multiple experiments conducted in the present work have been discussed as follows.
In Chapter 4, the results of conducting a series of experiments to design the CNN model-based CAC systems are given, and from the results of these experiments it can be concluded that the CAC system designed using the GoogLeNet CNN model yields an accuracy of 90.00% for the binary classification of chest radiographs, which is the best when compared to other CAC system designed using the AlexNet CNN model yielding 89.00% accuracy and the CAC system designed using the ResNet18 CNN model yielding 88.00% accuracy.
On the basis of the results of experiments conducted in Chapter 4, the CAC system designed using GoogLeNet CNN model performs the best; hence, it is used as a deep feature extractor in these chapters to design the CAC systems with different machine learning classifiers, namely ANFC-LH and PCA-SVM. From the results of the experiments carried out in Chapters 5 and 6, the CAC system designed using deep feature extraction by GoogLeNet and ANFC-LH yields the maximum accuracy of 93.00%.
Chapter 7 conducts a series of experiments to design the lightweight CNN model-based CAC systems, and from the results of the experiments conducted, it can be concluded that the CAC system designed using the lightweight MobileNetV2 CNN model yields 94.00% accuracy for the binary classification of chest radiographs, which is the best compared to other CAC systems designed using the lightweight SqueezeNet CNN model yielding 83.00% accuracy and the CAC system designed using the lightweight ShuffleNet CNN model yielding 92.00% accuracy.
From the results of experiments conducted in Chapter 7, the lightweight MobileNetV2 CNN model is used as a deep feature extractor to design a CAC system for the binary classification of chest radiographs. The results of the experiments conducted in Chapters 8 and 9 show that the CAC system designed using deep feature extraction by lightweight MobileNetV2 CNN model and ANFC-LH yields the same accuracy as the CAC system designed using deep feature extraction by lightweight MobileNetV2 CNN model and PCA-SM, which is of 95.00%.
Table 10.1 shows the comparative analysis of the different CAC systems designed in the present work for the classification of chest radiographs into Normal and Pneumonia.
Table 10.1
S. no. | Experiment | Network/classifier | Accuracy (%) | ICA_Normal (%) | ICA_Pneumonia (%) |
---|---|---|---|---|---|
1 | Designing end-to-end pretrained CNN-based CAC system for chest radiographs using AlexNet | AlexNet/softmax | 89.00 | 98.00 | 80.00 |
2 | Designing end-to-end pretrained CNN-based CAC system for chest radiographs using ResNet-18 | ResNet-18/softmax | 88.00 | 100.00 | 76.00 |
3 | Designing end-to-end pretrained CNN-based CAC system for chest radiographs using GoogLeNet | GoogLeNet/softmax | 90.00 | 96.00 | 84.00 |
4 | Designing end-to-end pretrained CNN-based CAC system for chest radiographs using decision fusion | Decision fusion | 91.00 | 100.00 | 82.00 |
5 | Designing hybrid CAC system for chest radiographs using deep feature extraction, GoogLeNet and ANFC-LH classifier | GoogLeNet/ANFC-LH | 93.00 | 96.00 | 90.00 |
6 | Designing hybrid CAC system for chest radiographs using deep feature extraction, GoogLeNet and PCA-SVM classifier | GoogLeNet/PCA-SVM | 91.00 | 96.00 | 86.00 |
7 | Designing lightweight end-to-end pretrained CNN-based CAC system for chest radiographs using SqueezeNet | SqueezeNet/softmax | 83.00 | 98.00 | 68.00 |
8 | Designing lightweight end-to-end pretrained CNN-based CAC system for chest radiographs using ShuffleNet | ShuffleNet/softmax | 92.00 | 100.00 | 84.00 |
9 | Designing lightweight end-to-end pretrained CNN-based CAC system for chest radiographs using MobileNetV2 | MobileNetV2/softmax | 94.00 | 98.00 | 90.00 |
10 | Designing lightweight end-to-end pretrained CNN-based CAC system for chest radiographs using decision fusion | Decision fusion | 95.00 | 100.00 | 90.00 |
11 | Designing hybrid CAC system for chest radiographs using deep feature extraction, lightweight MobileNetV2 and ANFC-LH classifier | MobileNetV2/ANFC-LH | 95.00 | 98.00 | 92.00 |
12 | Designing hybrid CAC system for chest radiographs using deep feature extraction, lightweight MobileNetV2 and PCA-SVM classifier | MobileNetV2/PCA-SVM | 95.00 | 100.00 | 90.00 |
ICA_Normal, individual class accuracy for Normal class; ICA_Pneumonia, individual class accuracy for Pneumonia class.
From the results presented in Table 10.1, it can be concluded that the CAC systems designed using the pretrained GoogLeNet CNN model yields an accuracy of 94.00%, which is the maximum when compared to other CAC systems designed using AlexNet CNN model and ResNet18 CNN model. Also the CAC system designed using deep feature extraction by GoogLeNet and ANFC-LH yields higher accuracy than the CAC system designed using deep feature extraction by GoogLeNet and PCA-SVM machine learning classifiers. Additionally, it yields an individual class accuracy (ICA) of Pneumonia that is the same as the CAC system designed using GoogLeNet CNN model.
The CAC system designed using deep feature extraction by lightweight MobileNetV2 CNN model and machine learning classifiers yields a higher accuracy (95.00%) as compared to the CAC system designed using lightweight MobileNetV2 CNN model.
For future work, other pretrained CNN models, such as VGGNet and ResNet-50, and other variants, such as DarkNet and NasNet, can be used to design CAC systems for the binary classification of chest radiographs. Similarly, deep feature extraction can be used on different layers of the CNN models, and various other machine learning-based algorithms can be used. Furthermore, hybrid CAC systems can be designed using different techniques of feature fusion, decision fusion, and a combination of handcrafted features with deep features extracted. These CAC systems could be designed for multiclass classification such as the three-class classification of chest radiographs into Normal, Pneumonia, and COVID-19 classes.
Fig. 10.1 shows a schematic representation of a decision fusion-based multiclass CAC system design. This can be designed by using multiple end-to-end pretrained CNN models as decision makers and then performing different types of decision fusion to achieve a final decision. These CAC systems can be designed for binary as well as multiclass classification of the input data.
Fig. 10.2 shows a schematic representation of a feature fusion-based hybrid multiclass CAC system design. This type of CAC system can be designed by performing the feature fusion of the deep features extracted using the CNN models and the handcrafted features. These extracted features can be fused using various techniques, and then machine learning-based classifiers can be used to perform the final classification, either binary or multiclass.
Fig. 10.3 shows a schematic representation of a feature fusion-based hybrid multiclass CAC system design. This type of hybrid CAC system can be designed by performing deep feature extraction from multiple pretrained CNN models and then applying different feature fusion strategies to ultimately perform classification using machine learning-based classifiers.
Fig. 10.4 shows a schematic representation of deep learning-based hierarchical multiclass CAC system design. This type of hierarchical CAC system can be designed by training an end-to-end CNN model to perform classification in the manner similar to a tree structure.
Fig. 10.5 shows a schematic representation of hybrid hierarchical CAC system design. Similar to the previously discussed CAC system, this hybrid hierarchical CAC system applies the machine learning-based classifiers to perform classification.
These CAC system designs previously discussed can be implemented not only with pretrained CNN models but also with self-designed CNN models, lightweight CNN models, and other variations of deep learning-based networks. These CAC systems can be applied for binary and multiclass classification of chest radiographs as well as for other imaging modalities and different tissue diseases.