A Comparison of Pre-Trained Models for Pneumonia Disease Prediction Using Chest Images

Aim : As viral diseases like Corona spread from one person to another, it has great impact on the public health system and socio-economic activities all over the world. Material and method : The only way to solve the spreading of this disease is early diagnosis of this disease. Statistics and Result: Deep learning algorithms were utilized in this study for comparative analysis of pre-trained models such as VGG16, MobileNetV2 for the detection of pneumonia using different hyper parameters such as batch-size, learning rate, epochs and so on. The proposed models that are MobileNetV2 and VGG16 attains better performance.


Introduction
The corona virus has been linked to cases of severe acute respiratory syndrome.The clinical symptoms of pneumonia infected patients are identical to the signs of illness caused by bacterial pneumonia (Lu et al., 2020;WHO, 2020).The rise in instances was brought on by direct contact between individuals via respiratory droplets (Chavez ET AL., 2020;Corman et al., 2020).Chest Xray images are frequently employed in the detection of lung illnesses (Guo et al., 2020;Lijens et al., 2017).due to its low cost and less radiation.In this study Deep learning architectures (Cheng et al., 2016;Lakshmanaprabu et al., 2019), such as VGG16 and MobileNet-V2 utilized and evaluated on chest X-ray image dataset.

VGG16 Architecture:
One of the Convolution Neural Network models is the VGG16 model, which has 16 layers.It has layers that are fully linked, fully pooled, max pooled, and convolutional.Fig. 1 mentions the VGG16 architectural block diagram.

Materials And Methods
Framework suggested for distinguishing pneumonia infection from X-ray images as follows.The framework for disease diagnosis is depicted in fig 4.

Dataset Description
The image dataset was considered from the Kaggle pneumonia X-ray dataset.It encompasses 3000 "X-ray images" of which 1500 are "pneumonia X-Ray images" and 1500 are "Normal X-ray images".The "normal X-ray images" and "pneumonia X-Ray images" are represented in fig 5a and fig 5b respectively.

Results and Discussion
In this study deep learning techniques were analyzed on the image dataset with performance parameters.Evaluating the Performance of the techniques based on the parameters like accuracy, precision, recall, F1 score and specificity.Confusion matrix is utilized for evaluating the metrics based on actual class and predicted class.The performance metrics are evaluated as follows: • Accuracy = "TP + TN" / "TP + TN + FP + FN".

Conclusion
The efficient detection of pneumonia infection due to Carona virus is possible by analyzing images.In this study "VGG16" and "MobileNetV2" were considered for the detection of the "pneumonia X-ray images" from "pneumonia X-ray images".The effectiveness has been improved with data preprocessing, augmentation and transfer learning techniques by enhancing visual quality of the image.Performance of these models were evaluated based on Precision, recall, F1score, and accuracy.

Fig. 3 .
Fig.3.Overall architecture of the MobileNetV2 Related Work Moutounet-Cartan et.al.(2020) investigated deep neural network models of the VGG16 and VGG19 for the detection of pneumonia infection from chest X-ray images.Khan et.al. (2020) proposed coronet model for accurate diagnosis pneumonia patients by CT images.Wu et al. (2020) established a model on chest CT images and it shows better performance.Deep learning frameworks have been widely used for the detection of pneumonia throughout the past few years (Bhandary et al., 2020; Kermany et al., 2018).Hashmi et al. introduced a novel approach based on a weighted classifier, which combines the weighted predictions from the state-of-the-art deep learning models such asResNet18, Xception, InceptionV3, DenseNet121, and MobileNetV3 (Hasmi et al., 2020).employed deep transfer learning models GoogLeNet, ResNet-18, and DenseNet-121(Kundu et al., 2021).Many

Fig
Fig 5a.Normal X-ray images Fig 5b.pneumonia X-ray images Data Pre-processing Images are pre-processed to improve image quality and to reduce the noise.It includes image resizing, image normalization, data augmentation, data splitting and transfer learning techniques were utilized for enhancing the quality of image and for enhanced performance.

Fig
Fig 6.Confusion Matrix VGG16 and MobileNetV2 models are analyzed based on their accuracy, specificity, and sensitivity for epoch 50, 75 and 100 with learning rate 1 and batch size of 8 and training and testing ratio is 80:20 and represented in table 5. VGG 16 analysis is represented graphically in fig-20, fig-21 and fig-22.Whereas MobileNetV2 analysis is represented graphically in fig-23, fig-24 and fig-25.

Fig
Fig 7. Confusion Matrices of VGG16 with Epochs 50

Table 2 :
MobileNetV2 AnalysisMobileNetV2 indicates greatest "accuracy" of 96%, "sensitivity" of 96%, "specificity" of 97% and average of recall and f1score of 96% when it is compared with the VGG16 model based on table 1 and table 2. Sample confusion matrix represented in fig 6.
Confusion Matrices for the VGG16 model with epochs 50, 75 and 100 were represented in fig-7, fig-8 and fig-9.Similarly, confusion Matrices for the MobileNetV2 model with epochs 50, 75 and 100 were represented in fig-10, fig-11 and fig-12 respectively.The confusion matrix values for VGG16 and MobileNetV2 are represented in table 3.

Table 3 : Confusion matrix values of VGG 16 and MobileNetV2 Analysis VGG 16 MobileNetV2 Epoch-50 Epoch- 75 Epoch- 100 Epoch- 50 Epoch-75 Epoch-100
The loss and accuracy metrics are also used for the performance evaluation of models.High accuracy and minimum loss are required for greater performance that is represented in fig-19.The "loss" and "accuracy" values for both "training" and "validation" for VGG 16 and MobileNetV2 models for epoch 50, 75 and 100 are evaluated and represented in table 4.

Table 4 :
Loss and Accuracy values of VGG 16 and MobileNetV2 Analysis

Table 5 :
Comparison of VGG 16 and MobileNetV2 Analysis