D-Cov19Net: A DNN based COVID-19 detection system using lung sound

Abstract

The limitations of proper detectors for COVID-19 for the proliferating number of patients provoked us to build an auto-diagnosis system to detect COVID-19 infection using only one parameter. Our designed model is based on Deep Convolution Neural Network and considers lung/respiratory sound as the deterministic input for our approach. 'D-Cov19Net' has been trained with 23,592 recordings, begetting an AUC of 0.972 and sensitivity of 0.983 after 100 epochs. The model can be of immense utility in biomedical technology due to its significant accuracy, simplicity, user convenience, feasibility, and faster detection while maintaining social distancing.

1. Introduction

COVID-19 pandemic has taken its toll on millions of lives worldwide since its unfortunate outbreak [1], [2], [3]. The worsening situation has led medical, engineering, and related researchers to brainstorm ideas to stop the rapidly increasing death counts with each passing day. As a need of the hour, various effective and ingenious methods to detect the presence of the deadly virus in the human body at the earliest possible stage have also been proposed and devised.

One widely used, 'gold standard,' reliable and effective laboratory test for COVID-19 detection is Reverse Transcription Polymerase Chain Reaction (RT-PCR) [4], [5]. But, due to the rapidly increasing demand, the proper kit required to perform RT-PCR worldwide, alongside constraints like timing for specimen collection and the inconvenience caused due to the generation of false-negative results [5], [6], the quest for another effective yet more reliable method went on. Another pitfall of RT-PCR lies in the fact that abnormalities, later recorded as COVID-19 infection, were observed in Chest Radiograph or Chest Tomography scans but were reported to be negative in some cases RT-PCR [7], [8], [9], [10], [11]. Ghoshal and Tucker [12] proposed a neural network with the utility of Bayesian Convolutional Neural Networks (BCNN), with a significantly large dataset. A lightweight Deep Neural Network (DNN) has been proposed by Li et al. [13], bearing accuracy of 88 %. Xu et al. [14] devised a 3-D Deep Learning model after data classification, wherein the accuracy level was 87.6 %. Shan et al. [15] devised a method for segmenting and quantifying the infected regions. J. D. Arias et al. [16] opted for detection using deep learning with chest X-Ray as a parameter. Bai et al. [17] devised a method to distinguish the COVID-19 infection from pneumonia by using an enormous dataset comprising 132 583 CT slices and a bearing accuracy of 96 %.

Despite remarkable accuracy and efficiency, these methods were insufficient to ensure clinically acceptable detection [7] for the increasing number of patients. Moreover, the feasibility and user convenience were yet to be improved in some of the cases. Therefore, the thrust for finding a yet more affordable and feasible tool that could help in early detection of COVID-19 infection even in the areas where biomedical testing is limited or error-prone was in process.

Keeping the constraints in mind, we have developed our architecture, named 'D-Cov19Net', to detect the infection, using lung sounds as the parameter for detection. The model has been developed with a deep learning approach and is presently ahead of the pathological tests in terms of convenience and accuracy. Moreover, the use of lung sound as the sole parameter for detection makes our approach stand apart from the developed methods of detecting COVID-19 with the aid of Artificial Intelligence. D-Cov19Net considers lung sounds the input time-series signal for the model and converted to a 2D equivalent signal. Then it is processed by the separable convolution layers of the network and has been embellished with the proper usage of ReLu activation [18] and Softmax functions [19] to bring forth an accuracy of 97.22 %, which can certainly play a significant role in the detection of the infection. Moreover, the sole purpose of our architecture does not remain with detection solely, but also with the auto-diagnosis report in the form of a spectrogram and its analysis to provide the user with a coherent picture of the severity of the disease in the user under consideration.

The proposed model can function efficiently and bear significant accuracy, reducing the time complexity without any human intervention in the midst. Apart from these, it performs on a non-contact basis, thus ensuring the maintenance of the COVID-19 norms of social distancing. The overall model described in this paper may not bear the highest possible accuracy, but it can be surely said that it is much more feasible and cost-effective as it treats lung sound as the sole physiological parameter to the inserted by the user. The rest of the task is performed by an auto-diagnosis system, which can generate a report to ensure the detection of the infection at an early stage, which is a crucial parameter for the patient's recovery. Moreover, the development of the designed tool will yield more academically useful applications that could facilitate biomedical research and engineering domain. This method can serve well in areas especially where pathological tests are difficult to carry for such an enormous number of people. The architecture can be developed further to detect other diseases, especially respiratory diseases, incorporating required changes. All of the points have been summarised in Fig. 1.

Fig. 1.

The reasons and benefits for building the model.

The materials and methodology used in our designed method have been elaborated in Section II. Section III highlights the results and discussions. Finally, the paper is concluded in Section IV. The references have been cited in section V.

2. Materials and methods

2.1. Dataset collection

Primarily the data has been collected from Kaggle [20] and others by manual collection from different sources [21], [22], [23], [24]. The origin of the Kaggle dataset can be traced from Portugal and Greece. 920 recordings have been taken into account. A total of 5.5 h of recording has been collected from 126 patients, varying in duration from 10 s to 90 s. It can be further classified into 506 crackles and wheezes, 886 wheezes, and 1864 crackles, thus accounting for 6898 respiratory cycles and ultimately 920 recordings. The dataset consisted of both clean and noisy recordings to bring forth the conditions that occur in reality. The dataset has been collected from children, adults, and elderly persons. The respiratory sounds have been captured using the digital stethoscope and other techniques. The positions from where the recordings have been done include Trachea (Tc), Anterior left (Al), Anterior right (Ar), Posterior left (Pl), Posterior right (Pr), Lateral left (Ll), and Lateral right (Lr). These data are primarily used for the 'normal' and 'others' category in our dataset.

The ERS dataset [21] consists of 20 different case recordings of auscultated lung sounds. Among these 20 recordings, there are 11 wheezes, five crackles, two recordings of patients with pneumonia and lung cancer, and two recordings of patients with pleural bleeding and pleural effusion. The Medzcool dataset [22] is a sample dataset that consists of auscultated lung sounds of COVID-19. These consist of wheezes, fine crackles, and coarse crackles, making a total of 4 samples, and each sample varies in duration from 5 s to 10 s. The EMT dataset [23] is also a sample dataset of auscultated lung sound, which includes bronchial breath, coarse crackles, fine crackles, diminished breath sound, expiratory wheeze, pleural rub, rhonchi, stridor, vesicular breathing sound, which makes a total of 9 samples and each sample has a duration of 10–20 s. These data are used for the serious lung diseases in the 'others' category and the confirmed cases of the 'Covid-19' category.

The SARS-Cov-2 dataset [24] consists of auscultated lung sounds of 30 different COVID-affected patients. The auscultation process includes a Bluetooth digital stethoscope (Stemoscope) used on six bilateral pulmonary fields to record the required sound. Two are at the basal field, two at the middle field, and two at the upper field, both on anterior and posterior position, respectively. For a detailed recording of auscultated sound, a frequency range of 20 Hz to 1 kHz was used. Every subject was instructed to breathe deeply for the 30 s, including 2 s of constant inspiration and 2 s of constant expiration. During the measurement process, subjects were instructed to sit upright or have a bed elevation of 45–90 degrees in case of the forbidden patients. These data are primarily used for the 'Covid-19' category in which all the subjects are confirmed cases of COVID-19 disease.

Finally, data augmentation was applied to the existing real data to enhance our dataset. For augmentation, time- stretching techniques at multiple levels along with slight pitch variation were used. In case of time stretching, we slowed down and sped up the audio signal by a factor of t = 0.5–1.5 (0.5, 0.75, 1.25, 1.5 are the respective rates which were used), and for pitch variation, we pitched up and down the audio signal by a factor of p = −1 to 1 (−1, −0.5, 0.5, 1 are the respective factors which were used) semitone which was exported as a single data and then added to the original dataset. Thus, the final data in our dataset increased from 983 to 23,592 which is 24 times the original dataset, thus increasing varience in the data set yet retaining the features of a specific sample data.

2.2. D-Cov19Net: The proposed model

The D-Cov19Net has been designed based on tiled Convolution Neural networks [25]. The data obtained is normalized, as one of the effective pre-processing techniques before feeding into the architecture, to beget the required output, with the least possible time complexity [26].

The Convolution function can be generally stated as [27],

$$(f*g)(t)\triangleq {\int }_{-\infty }^{\infty }f(\tau )g(t-\tau )d\tau \dots$$ (1)

Where, (τ) is a weighting function, and τ is the parameter defining the function (t). But, as our signal was digital in nature so the input values will be discrete, thus changing the formula to,

$$(\mathrm{f}*\mathrm{g})(\mathrm{t})=\Delta {\sum }_{\tau =-\infty }^{\infty }f(\tau )g(t-\tau )\dots$$ (2)

The 1D time-series signal has been converted to a 2D equivalent signal, resembling the kernel size of the input layer. Thus, if we consider our input to be 'I' and denote the kernel by 'K,' the equation can be mathematically stated as to where the symbols have their usual meanings [27],

$$(\mathrm{f}*\mathrm{g})(\mathrm{t})=\Delta {\sum }_{n}I(m,n)K(i-m,j-n)\dots$$ (3)

The respiratory sounds are to be recorded by the lung auscultation process with a digital stethoscope, and then it is converted into a two-dimensional form by Mel Spectrogram. The architecture uses the zero-padding technique.

[28] to ensure that feature learning efficiently takes place. For the sake of enhancing the efficiency of our proposed architecture and optimizing the excess computational power, the depth-wise separable design based on 2D Convolutional layers has been implied.

D-Cov19Net is designed to be inspired by the Xception architecture [29], with numerous alterations to beget more accurate results and make our proposed method more convenient and effective. In D-Cov19Net, each convolution layer has been optimized using batch normalization [30] and ReLU activation functions, except for the last convolution layer. For avoiding complexions, two 2D convolution layers (3 ×3) were stacked in the first layer to ensure lightness and proper functionality.

The first two layers have been optimized using batch normalization and ReLU activation. Batch normalization helps accelerate the learning rate and makes the network more stable through normalization of the input layer by re-centring and re-scaling the concerned data.

Our algorithm introduced three different depth-wise separable convolution sections with the input bearing kernel size of 128×128×1. After pre-processing, the dataset was configured to fit into the model, and batch normalization was included to prevent overfitting or underfitting and to have a stable learning process. ( Fig. 2).

Fig. 2.

The illustration of the data augmentation where S_i is the initial data.

Here, we consider J(V, K) as the loss function that we want to minimize, now, due to the utilization of the backpropagation function [31], we will find a tensor G with its elements. Next, we have computed the derivatives concerning the weights in the kernel. The mathematical function, which has been instrumental in making our architecture, can be stated mathematically as,

$$\frac{\partial }{\partial Ki,j,k,l}J(V,K)={\sum }_{m,n}{G}_{i,m,n}{V}_j,(m-1)\times s+k,(n-1)\times s+l\dots$$ (4)

To backpropagate the error further down, we have computed the gradient concerning V. Several separable layers have been introduced in the middle flow section, with eight stacked units to obtain a deeper network for more accurate results. The batch normalization has been used in a specific manner where each entry controlled separable layer is Batch normalized except the exiting layer. These beget more trainable parameters without overfitting the layers. The categorical cross-entropy loss function has been incorporated to calculate the loss of a sample from the three classes by computing the following sum:

$$\mathrm{Loss}=-{\sum }_{i=1}^{outputsize}{y}_i.log{\hat{y}}_i\dots$$ (5)

Where y _i denotes the i^th scalar value in the model output, y _i denotes the corresponding target value, and the output size denotes the number of scalar values in the model output.

In the exit flow, two dense or fully connected layers have been introduced, the penultimate layer being ReLU activated and the last dense layer is Softmax activated with three output classes "Covid," "Normal," "Others." After the pre-processing of the input signal is completed, converting the one-dimensional signal into two- dimensional form by spectrogram, which is further reshaped to 128 × 128, the model predicts the output class. The result so obtained acts as an auto-diagnosis report, thus detecting the COVID-19 infection in an individual.

The model has been trained with the pre-processed data incorporating the backpropagation algorithm and Adam optimization [32]. The model has been trained on 100 epochs.

The model's architecture has been summarized and presented conveniently using flow charts ( Fig. 3), and the model summary has also been included ( Table 1).

Fig. 3.

(from left) Brief illustration of the proposed network; (from top right): Illustration of Convolution section A, Illustration of Convolution section B, Illustration of Convolution section C.

Table 1.

Model Summary.

3. Results and discussions

Our model has been trained with the pre-processed data by backpropagation based on the Adam optimization algorithm on 100 epochs. 98.3 % of training accuracy and 97.2 % of validation accuracy were achieved after completion of training ( Fig. 4). The validation accuracy and training accuracy have been depicted graphically in Fig. 5.

Fig. 4.

Illustration of the proposed method.

Fig. 5.

Training loss and accuracy graph.

For instance, we used a real-time lung sound of a subject, 20 years old, which is recorded by lungs auscultation process through a digital stethoscope and fed it into the system for the feature extraction by Mel Spectrogram. After the extraction process is done, the audio is then fed into the model for the prediction. The output was 'Normal' with a prediction accuracy of 99.67 %. The output on the spectrogram has been depicted [ Fig. 6(a)]. Similarly, an input of a 1st stage COVID affected a person's lung sound, aged 32 years old was tested, and the output was 'COVID Positive' with a prediction accuracy of 98.79 % [Fig. 6(b)]. All the subjects used for testing are confirmed subjects as stated earlier in section II.

Fig. 6.

(a), 6(b): (from left) spectrogram of the normal lung sound and a COVID affected lung sound.

The main challenge was to give distinct predictions between the 'COVID Positive' and 'Others.' The 'Others' class encompasses the lung sounds of all the other diseases associated with the lungs, and pneumonia was among them. The reason behind using a normal Pneumonia-affected lung sound is that COVID-19 also causes pneumonia with certain deviations. To test this, we gave input on a Pneumonia patient's lung sound. The result came out to be as expected. It gave an output 'Others' with a prediction accuracy of 82.45 % [ Fig. 7].

Fig. 7.

Spectrogram of a Pneumonia affected lung sound.

The sounds used for testing the model were unique and were not present in the training dataset. To predict the overall testing accuracy of our model, we computed a confusion matrix based on four main parameters, which are the True Positive (T.P.), False Positive (F.P.), False Negative (F.N.), and True Negative (F.N.) on three different class each having 60 different test samples. From the computation of the confusion matrix on the test data set, we achieved a sensitivity of 98.33 % and a specificity of 96.667 %. Again, we computed the confusion matrix based on 'COVID Positive' and 'Others.' Here again, we achieved a sensitivity of 98.33 % and a specificity of 96.667 %. The results show that our proposed method can detect the true cases of COVID-19 positive with just a 1.67 % error rate, and on the other hand, our proposed method can detect the true cases of the COVID-19 negative with just a 3.333 % error rate. Thus, our proposed method proves to be highly sensitive and specific even if there are some critical cases. The overall accuracy is calculated through the confusion matrix [33], which is 97.223 %, considering the minimum amount of data that we used for the model training. The confusion matrix has been tabulated in Table 2.

Table 2.

Confusion Matrix for COVID-19 detection.

Class	Test 1 (COVID +ve)	Test 2 (COVID -ve)
COVID-19	TP= 59	FN= 1
Normal	FP= 2	TN= 58
Others	FP= 2	TN= 58

In addition to that, the analysis based on a histogram has also been performed [34]. The histogram plotting for normal, pneumonia affected, and COVID-19 affected patients have been plotted respectively in Fig. 8(a,b,c). The frequency data of each data point is averaged from all the samples for a specific category, and the resulted data is visualized by histogram. The histogram shows the average dominating frequencies with their respective minimum threshold value. From the figure, we can conclude that, though the histogram plotted for COVID-19 infected case and pneumonia case resemble, certain differences mark the distinction. Thus, all the results obtained so far indicates our method of detection to be efficient and effective.

Fig. 8.

(a,b,c): (from top left) Normal, pneumonia affected, COVID-19 affected.

Several methods for detecting COVID-19 have been devised in recent times using A.I. A comparison table has been plotted ( Table 3) to compare the accuracy for all the proposed methods using deep learning. It indicates our proposed method to be accurate than the majority of other proposed methods. Moreover, we notice that the detection accuracy of the infection is higher when using speech signals as input. But, we designed this model to be more efficient as it determines the infection in an individual directly by accounting for the acoustic parameters of the lung without increasing complexity.

Table 3.

Accuracy comparison of different COVID test.

Sl. No.	The method used for detection	Medical Parameter used for detection	Approximate accuracy (in percentage)
1	A.I. (Deep Learning)	C.T. Scan [35], [36]	90.8
2	A.I. (Deep Learning)	Chest X-Ray[16]	91.5
3	A.I. (Deep Learning)	Speech/cough recording [37], [38]	98
4	A.I. (Our proposed method)	Lung sound	97.22

To determine the accuracy and account for the validation of the model, we have plotted a ROC curve, and the sensitivity and AUC have been visualized from there [39]. The ROC curve ( Fig. 9) is plotted following the confusion matrix depicted above (Table 3).

Fig. 9.

ROC Curve.

This result indicates that our model can classify amidst the mentioned classes properly and the True Positive Rate being significantly higher depicts the sign of reliability on the outcome of our model. As a result, our suggested technique can detect and classify COVID-19 disease with excellent specificity and sensitivity.

4. Conclusion

The rapid spread of COVID-19 worldwide and the increasing number of deaths require urgent actions from all sectors. The key factors behind the rapid spread by continuously changing its phase of the COVID-19 pandemic, such as scarcity of testing and its cost and time-consuming procedure, prompt us to take some urgent action. This study reveals an all-over deployable deep-learning-based preliminary diagnosis tool for COVID-19 using lung sound samples.

The innovation and utility of this architecture lie in the fact that it outdoes different existing architectures in terms of efficiency. Another note is that the full implementation has been done successfully with a standard- sized dataset with remarkable accuracy. Moreover, this test for COVID detection is contamination-free, i.e., there will be no false prediction based on the contamination factor. To acquire a proper lung sound sample, the recording should be done in a controlled environment to eliminate the unwanted noise frequencies, resulting in an accurate prediction. And, the total proposed model can be used without any further modification in the areas of medical sciences to test for COVID infection efficiently without violating any preventive norms.

The main challenges included distinguishing lung sounds from a pneumonia patient and a COVID-19 infected patient. But, our model proved its efficiency in overcoming the challenge. Moreover, a huge amount of data is not required to train the model.

This architecture, with minimal modifications, can be used to function in the same way for some diseases which bear similar symptoms. Even though the results brought forth by the model seem promising, it can always be considered safe to get the analysis verified by some medical personnel, thus ensuring its clinical acceptability. With the involvement of the medical community and required modifications, the scope for building a clinically acceptable model can be accounted for using our proposed work. Our study and proposed model can also act as a building block of various other researches to detect lung diseases with the aid of Artificial Intelligence. With time and further development, this method can turn out to be of immense use in the medical domain for detection of crucial diseases in a glimpse, with minimal effort, that could play a pivotal role in saving the lives of many due to detection within the expected timeframe when medical science will still have something to do and cure the patient, instead of giving up just because, detection of the disease took place after the threshold of time, to save a precious life.

Thus, to conclude, all the graphs and illustrations have significantly highlighted our model to be legit with a desired level of accuracy. To account for the validation of our model, the ROC curve has been plotted, and the AUC and sensitivity have been calculated after that. With a significant outcome, i.e., an AUC of 0.972 and a sensitivity of 0.983, it can be concluded that our model is reliable and academically useful.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Biographies

Anilesh Dey was born in West Bengal, India, in 1977. He received a B.E in Electronics from Nagpur University and M.Tech. (Gold-Medallist) in Instrumentation and Control Engineering from Calcutta University and received PhD. from Jadavpur University. He is working as an Associate Professor and is presently the Head of the Department of Electronics and Communication Engineering at the Narula Institute of Technology, Agarpara, Kolkata. He has been the author and co-author of more than 70 scientific papers in international/national journals and proceedings of the conferences with reviewing committee. He has conducted several research works in the domain of biomedical engineering.

Sukanya Chatterjee was born in West Bengal, India, in 2001. She is pursuing her B.Tech graduation in Electronics and Communication Engineering from Narula Institute of Technology, Agarpara, Kolkata

Jishnu Roychowdhury was born in West Bengal, India, in 2000. He is pursuing his B.Tech graduation in Electronics and Communication Engineering from Narula Institute of Technology, Agarpara, Kolkata