The new coronavirus (COVID-19), declared by the World Health Organization (WHO) as a pandemic, has infected more than 1 million people and killed more than 50 thousand. This disease is caused a by severe acute respiratory syndrome coronavirus-2 (SARS-COV-2). All the patients are being observed with visual characteristics in the chest x-rays.
The dataset constructed for this study are composed of 5.683 images of patients diagnosed with coronavirus and healthy people. We apply the concept of transfer learning for this task. We use a convolutional neural network (CNN) architecture trained on VGG-16, adapt them to behave as feature extractors for the x-ray images.
Deep LearningΜany clinical and radiological studies have been performed based on deep learning techniques. For instance, many convolutional neural networks known as CNNs, recursive networks, transfer learning models, etc. have been implemented to analyse automatically radiological disease features used a convolutional neural network. The development of convolution neural networks layers has allowed a great perform in ability to classify images and detect objects in images. So, we used the VGG-16 pretrained model for our study.
In order to classify the X-ray images, we perform three classifications steps. These are: i) a model training ii) a model testing and iii) a repetition of two above steps. The dataset includes images of 1000 classes, and is split into three sets: training (70%), validation (15%), and testing (15%). Our test set contains 852 images of two labeled categories. After that, the test images are given as input to the pretrained CNN, where all the parameters are already optimised. Then the CNN extracts the images features and classifies them into appropriate classes using soft-max classifiers.
By creating decision support systems (DSS) we want to deploy very efficient model that may potentially improve diagnostics and predictive accuracy. These DSSs will detect abnormalities and so they will be classified images based on normal and covid labels. Models just for the record are our hope to accelerate the biomedical discoveries and provide clinicians with new tools. This process can use Artificial Intelligence and Machine Learning approaches. The CNN architecture contains a number of layers. These convolutional layers are responsible for extracting features from the input images using several convolutional filters. Many times, in DSS deployment is been utilised an untrained model which requires a long training process to optimize the weights. Many models have developed by many data scientists, such as OptCoNet, ResNet, Chest, CovidX-Net, ALexNet, etc.
Evaluation MetricsIn this section we present the results of image classification results achieved by described ConvNet architectures. After the training we are going to visualize all metrics and proceed to some significant notices about our model. Generally, metrics as accuracy, F1-score and false positive rate (FPR) are important for the evaluation of DDS.
True positive (TP) indicates the number of correctly classified 5678patients, true negative (TN) informs the number of healthy patients that were correctly classified. False negative (FN) corresponds to the number of occasions that were misclassified a health patient and false positive (FP) points out the number of times that model was wrongly classified healthy patients. All the equations about metrics are presented below (1)-(4), respectively.
Fig.4-5. show the metrics we describe in previous sections. Analysing figures, we reached an overall accuracy of approximately 98% at the same time with a validation accuracy of 100% and FPR of 2.1%. For further analysis of the results, our model has a 98% specificity and 97% sensitivity. That means a significantly increased percentage of our images can be correctly classified by our algorithm.
In this study, the main objective was to classify the X-ray images of people suffering from COVID pneumonia and normal people. Convolution neural networks algorithms have been used. After all the necessary preprocessing we used the VGG-16 architecture. After building and executing the whole process, we observed that great perform of the model. Searching for comparison studies we summed up that is the best among all the existing ones. So, we can step forward and examinate a potential classification of covid versus bacterial pneumonia versus normal images. Essentially, the accurate differentiation of covid and bacterial pneumonia is the endgame of our case study. But the findings of such research are still in the early stages.
Comments
Please log in or sign up to comment.