Keywords

1 Introduction

The ubiquitous COVID-19 pandemic is the distinguishing tragedy of the present century thus far. To combat this battle, clinicians and radiologists work together and invented a couple of different diagnostic tools. COVID-19 reportedly started in Wuhan City of China in December 2019 [1,2,3]. According to the Centers for Disease Control and Prevention (CDC), the total number of cases and deaths as of 8th May 2022 is 517 million and 6.25 million, respectively worldwide. That is why this research has a significant value from the social perspective as it helps faster the processes and automate the tasks rapidly and build more confidence in the medical fraternity. The numbers continue to increase with the spread of the Omicron coronavirus variant (B.1.1.529). COVID-19 is a disease that initially affects the respiratory organs, like the lungs [1]. As it is a highly transmittable viral infection without a cure, people are dying all over the world and many severely suffering patients do not find hospital accommodation due to overloading with the existing COVID-19 patients. The Critical Care facility is also overloaded and the situation is overwhelming.

The diagnosis of COVID-19 is one of the crucial aspects of the fight against this disease. In the early days of the pandemic, the detection was very limited and error-prone due to a lack of understanding of the virus genome, but researchers came up with the Reverse Transcription Polymerase Chain Reaction (RT-PCR) [4] pathological method, which is still a gold standard for the detection of COVID-19. However, many challenges still remain in the COVID-19 detection research. The RT-PCR sensitivity rate is 60%–70%, which is very low. Rapid antigen testing is also a fast method to detect COVID-19 patients but due to the different types of variants coming up every now and then it is also not so effective as lots of True negatives coming in the test due to the mutation of the virus genome.

Recently, many automated machine learning approaches have been introduced in the literature for Covid-19 detection such as Chest CXR, Chest CT, etc.

In this paper, we present a very efficient deep learning model for analyzing CT images.

2 Literature Review

As the situation in early 2020 had already started deteriorating and new cases were growing exponentially, The World Health Organization (WHO), declared it a pandemic on 11th March 2020 [5]. WHO distinguishes the variants of coronavirus into two categories, one is a ‘variant of interest’ and the other one is a ‘variant of concern’. According to the transmissibility and detrimental effects, the variant of concern is more dangerous compared to the variant of interest.

He, X. et al. [6] present a deep learning diagnosis method using chest CT images. As their work was reported in early April 2020, their COVID-19 chest CT image dataset was limited. Zhang, K. et al. [7] proposed an AI System for diagnosis, quantitative measurements, and prognosis of COVID-19 pneumonia using computed tomography. As their research was published in early June of 2020 they used the idea of image slices to increase the size of sample data as only 752 COVID-19 positive patient data was included in their work. Gunraj, H. et al. [8] proposed a tailored deep convolutional neural network (CNN) using chest CT images. They generated a dataset of 104,009 chest CT slices across 1,489 patient cases. Harmon, S. A. et al. [9] present a deep neural network using a multinational dataset. M. Khushi et al. [10] and Maillo, Triguero and Herrera [11] show that having limited positive samples in a dataset could lead to data inconsistency and data imbalance issues. Panwar, H. et al. [12] proposes a grad-CAM-based color visualization in their deep learning approach using CT images.

3 Our Proposed Methodology

Deep Convolutional Neural Network (CNN) is one of the most widely used neural networks for image classification problems [13]. CNN in a deep learning setting has been widely applied in the literature, for example, [14,15,16,17,18]. With the availability of different deep learning networks such as ResNet18, Res-Net50 [19], AlexNet [20], DenseNet [21], VGG16 [22], EfficientNet [24], etc., one can design many automated machine learning models that can detect COVID-19 from CXR, CT images, or a combination of both. In our approach, we designed our own classification layer but leveraged the advantages of EfficientNet in multi-feature generation layers through the transfer learning technique. It helps to generate lower-level and higher-level features in a rapid way. Iterative multiple training helps to optimize parameters and hyperparameters with rapid succession due to this integrated deep multi-layer approach.

3.1 Dataset Description and Preprocessing

The data consisted of chest CT images divided into two categories. COVID-19 infected patients and non-COVID-19 patients. We used multiple sources for data acquisition [9, 12,13,14,15] as the opensource COVID-19 data increases over time. These images are labeled by licensed radiologists and freely available in open-source platforms like GITHUB repositories and Kaggle platforms. Once we obtained the CT images, our next step was to preprocess those images with proper annotation and labeling, including normalization and augmentation using machine learning models. We use Pytorch vison models [23], to flip, rotate and include those images to increase the size of the dataset, which basically helps the deep learning model to be more robust and accurate. The deep neural network models derives feature information in a very different way they human radiologist visually inspect the images. Initially, some sort of histogram equalizing and contrast-enhancing.

Fig. 1.
figure 1

Chest CT images after preprocessing.

Full size image

Our proposed model used a dataset of 8054 real patient CT scans, of which 5427 had COVID-19 and 4223 were Non-COVID-19 patient images. The dataset is further divided into train, validation, and test sets with the ratio of 7:2:1. Figure 1 shows the chest CT images after preprocessing. Our dataset is very well balanced thus does not incurred data imbalance issues. Data imbalance is a very common problem, especially happened in medical data and image analysis if the model used fewer number of positive sample cases than the negative sample cases.

3.2 Methodology

The Fig. 2. It shows the block diagram representation of our proposed model. EfficientNet [24], pretrained weightages helps to reduce the time of our training process as due to transfer learning technique of EfficientNet we used the hyperparameter values in our model.

Fig. 2.
figure 2

Abstract Block Diagram Representation of our Proposed Model

Full size image

After preprocessing the images we used EfficientNet [24], a pretrained Model, for the feature generation tasks and then we add a deep classification layer to classify the CT images into COVID-19 position images and COVID-19 negative images, i.e. Non-COVID-19 images. After execution of the feature generation tasks, the total number of parameters of our dataset is 18,830,011, out of which Trainable parameters are 1,281,395 and Non-trainable parameters are 17,548,616.

The architecture of EfficientNet baseline Neural Network is shown in Fig. 3.

Fig. 3.
figure 3

EfficientNet B0 Baseline neural network Architecture

Full size image

EfficientNet is a very efficient high-performance convolutional neural network architecture and scaling method that uniformly scales all dimensions of depth, width and resolution using a compound coefficient. Where benchmark deep neural models arbitrarily scale these factors, the EfficientNet scaling method uniformly scales network width, depth, and resolution with a set of fixed scaling coefficients. The effectiveness of model scaling is based on the baseline network which is playing a key role in high performance. The baseline network builds up with a deep neural architecture search using the AutoML MNAS framework [25], which optimizes both accuracy and efficiency. After leveraging the potential of transfer learning by using the deep learning model EfficientNet in the feature generation architecture we design our deep learning classification architecture. The elements of the architecture is given in the Fig. 4. We use piecewise linear function rectified linear unit in our model as an activation function instead of Softmax or other activation function as we are interested in getting probabilistic value from 0 to 1.

Fig. 4.
figure 4

Classification Architecture of our Deep Learning Model

Full size image

The classification layers provide the percentage of the score for COVID-19 patients and Non-COVID-19 patients.

4 Results

The proposed deep learning model had an average detection accuracy of 91.96% on the test CT images. Sensitivity is 92.24, Specificity is 93.01 and F1 Score is 0.90. Figure 5 shows the Loss Curve. We used three fold cross validation technique and after the 20 epoch as the values converge to the optimum values so no further training of the model needed, We heuristically checked this is the optimum level of the loss in our experiment while running with many epoch heuristically to understand the optimum level. Hence, we only included the optimum value figures upto 20 epoch. From the graph, in Fig. 5a, it is evident the loss decreased over the epoch. Figure 5.b shows the accuracy curve. From the graph, it is evident that the accuracy improves over the epoch.

Fig. 5.
figure 5

a) Training and Validation Loss b) Training Accuracy Vs Validation Accuracy

Full size image

Figure 6a, b, c are the results coming from an unknown test sample and then matching their value for the predicted score obtained using our proposed model. It is clearly evident from the Fig. 6 that the predicted score clearly confirmed the type of the class with very high accuracy.

Fig. 6.
figure 6

Predicted score matching with the ground truth values for different test images.

Full size image

The following Table 1 presents the comparison of the proposed model with other existing models using our dataset with the model they described in their papers.

Table 1. Comparison Analysis with Existing Models
Full size table

5 Future Work and Conclusion

The Deep learning Model is not intended to replace the RT-PCR or other bio-chemist methods rather it would help to give extra confidence to the doctors to detect COVID-19 with more accuracy, hence helps to reduce wrong diagnosis. Many earlier models reported in the literature utilized a limited number of COVID-19 positive images due to the lack of availability of such data during the primary phases of the pandemic. In the present work, 8054 CT images were used with balanced positive and negative cases. The results show that this work which is based on EfficientNet is capable of detecting COVID-19 based on CT images with high accuracy. This adds to the growing body of evidence that deep learning techniques may be used in a clinical setting with confidence for detecting COVID-19. One of the recommendations of WHO is to increase testing for COVID-19. Usage of deep learning-based methods would bring down the cost and accelerate the diagnosis process which will enable rapid testing on a global scale. In our future work, we would investigate vision transformer models with this dataset. Transformer architecture performs very well in Natural Language Processing (NLP) tasks and is widely used in NLP domains. But recently many research works comes up with vision applications as well, though Deep Convolutional Neural Network is mostly preferred by most of the researchers in computer vision applications, especially in image processing domains, object detection domains, etc. We also like to extend our work on our classification layers with more variety of activation functions and loss functions.