Automated diagnosis of COVID-19 using radiological modalities and Artificial Intelligence functionalities: A retrospective study based on chest HRCT database

1. Introduction

With its emergence in late 2019, the persisting pandemic, namely the novel coronavirus (COVID-19), has drastically affected people across the globe. The outbreak of this deadly disease was caused by the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). The rapid transmission of this disease from human to human is phylogenetically close to bat SARS-like coronavirus with a separate monophyletic group [1]. The SARS-CoV-2 virus is a member of the human coronavirus (HCoV) family that affects the lower respiratory tract. Due to the high number of confirmed COVID-19 cases and no treatment for the infection, safety precautions are implemented worldwide, such as social isolation, the use of a mask to prevent the virus from entering the respiratory system, quarantine, and other containment measures, which are able to lower morbidity and mortality among highly susceptible individuals [2]. Still, studies reveal that, unlike the SARS-CoV and MERS-CoV viruses, the mortality rate is lower in the case of COVID-19 [3]. SARS-CoV-2 is largely spread through droplet inhalation, indirectly through contact with infected fomites, and through airborne inhalation of bioaerosols that are suspended in the air. The symptoms of COVID-19 are inconsistent, but common symptoms include fever, difficulty in breathing, cough, fatigue, loss of smell, taste, and headache [4]. These symptoms may occur between one to fourteen days if exposed to the virus. An infected person may experience a few of these symptoms, and it may last for a minimum of seven days. If the severity of symptoms continues for more than seven days, then there are chances that the patient may include symptoms like dyspnea, hypoxia, and the involvement of 50% of the lung. Considering the impact factors and statistics, studies disclose that COVID-19 is a severe case of pneumonia that affects the lungs. Such symptomatic patients are easy to identify and take corrective measures for. But the situation worsens when the patient is asymptomatic COVID-19 positive. Even though they often go unnoticed, these people may be contagious and help SARS-CoV-2 spread to healthy people. Despite the fact that such patients are responsible for fewer secondary infections than symptomatic cases, asymptomatic COVID-19 subjects can still spread SARS-CoV-2 to other people. It is noteworthy that asymptomatic COVID-19 cases have viral loads that are equal to or higher than symptomatic cases [5]. Asymptomatology and high viremia can thus coexist and represent a significant risk factor for the spread of COVID-19 infection [6]. Because they do not exhibit symptoms that would make one wonder if they have COVID-19. Several variants of SARS-CoV-2 have been brought into focus by the World Health Organization (WHO), designated as the Alpha, Beta, Gamma, Delta, and Omicron variants [7]. With their potential for rising transmission and virulence, these variants have contributed to the persistence of this pandemic. The SARS-CoV-2 Delta strain was discovered in India in late 2020 and has since been discovered in almost 60 other countries [5], [6]. A higher rate of transmission is possible in Delta compared to other SARS-CoV-2 variations. Testing and the infected patients’ record study revealed that the Delta variation may result in more serious lung deformities, leading to death in extreme cases. While mass vaccination campaigns are currently just getting started, several medications have shown in vitro activity against SARS-CoV-2 or potential clinical benefits. Early identification and appropriate treatment of immunologic complications can reduce morbidity and mortality in patients with COVID-19 infection. The COVID-19 vaccination was first administered in India on January 16, 2021. India had provided more than 2.04 billion doses of currently recognized vaccinations as of August 29, 2022, including the first, second, and booster doses. In India, 86.87% of the eligible population (12+) is fully vaccinated, and 94.48% of the eligible population (12+) has received at least one shot [https://vaccinate-india.in/dashboard].

It is also true that contaminated vaccine recipients may conveniently show only moderate symptoms, stay asymptomatic, or otherwise go unreported. When significant viral loads are proven, vaccine recipients may spread the infection even more subtly. This is due to the possibility that participants in the COVID-19 mass vaccination campaign would neglect their need to maintain social distance. Therefore, it is crucial to consistently follow hygiene practices including hand washing, face mask use, and other public health safety precautions, such as social isolation. Additionally, it is important to promote the proper use of personal protection equipment, such as surgical masks and filtering face pieces, in high-risk settings, including crowded areas, public transportation, and indoor establishments like schools or hospitals. Additionally, it is important to improve occupational health safety practices, such as health surveillance, screening, testing mainly self-testing with antigen tests, and contact tracing activities [4], [5], [6], [7], [8], [9], [10].

The reverse transcription-polymerase chain reaction (RT-PCR) is considered the primary diagnostic test for COVID-19, which can analyze thousands of samples in a single day and has a testing sensitivity of 95%. Thus, the RT-PCR technique is regarded as the gold standard for both qualitative and quantitative viral nucleic acid detection. The lesser sensitivity in the RT-PCR method demonstrates various analytical issues, including human mistakes, testing outside the diagnostic window, active viral recombination, and insufficiently validated assays, which compromise the diagnostic accuracy [2], [11], [12]. The viral loads in throat swabs are most substantial when the virus first manifests, and the virus may start to shed two to three days prior to the beginning of symptoms, making presymptomatic or asymptomatic transmission easier. Therefore, instead of considering it as a diagnostic standard, we may consider RT-PCR as the primary detection measure. As lung involvement is a part of coronavirus infection, the physicians suggest chest computer tomography (CT) as a mandate for the proper diagnosis and prognosis of COVID-19 in its early detection due to its high sensitivity (60%–98%), of a viral lung [11], [12], [13]. Although chest CT imaging is not considered a standard screening test protocol for COVID-19, it is beneficial for people with mild symptoms or asymptomatic ones or those with a negative RT-PCR experiencing mild symptoms or chest anomalies, or unexplained lung pathology. Nevertheless, chest CT also, proves to be helpful for recovered cases of COVID-19 (post-COVID). Other imaging techniques like X-ray and ultrasound also help to evaluate the disease progression, but diagnosis with chest CT is preferred due to its three-dimensional pulmonary view and versatility [14].

Decision making with the detection of chest CT is typically based on several parameters such as whether the patient is RT-PCR positive, whether the patient has post-COVID symptoms or any other disease with overlapping distortion in the chest/lung. Moreover, diagnosis with the help of chest CT sounds advantageous in both alternative diagnosis, prognosis, and figuring out complications in COVID-19, reducing the rate of severity and mortality in cases [15]. While monitoring these parameters in manual and traditional way, in many cases it is time consuming, tedious and repetitive. Machine Learning (ML) and Deep Learning (DL) tools may provide a helping hand to the physicians so as to increase the screening rate [16], [17], [18], [19], [19]. The automated ML and DL algorithms can be used in biomedical platforms and can be used as AI-based tools to design predictive models. These models, deployed on a computer system, learn to detect, classify, or diagnose the CT images. The AI-aided predictive diagnostic tool may be configured to avoid the tedious and repetitive task of manually evaluating each CT image of COVID patients. Whereas AI enables the computer to understand from a large set of databases, to identify or classify diseased cases, a proper methodology to configure the system is required to design and deploy it. In a previous work [20] on methods for COVID-19 detection, the authors have reviewed the current ongoing research with the reported databases. In that work, the authors compared various algorithms and schemes for classification, based on the reported findings. As a summary of that paper, in a nutshell, engineers need to use the signal processing on these CT scan images with ML and DL tools to help the physicians screen at a faster rate [21]. In this direction, the first criteria is to understand the CT scan images and their findings so that they can be incorporated into the algorithms for better identification and classification of diseased cases.

Studies have brought into focus several initial chest CT findings in COVID-19 positive cases, which include peripheral ground-glass opacity (GGO), consolidation, GGO with consolidation, and crazy pavings [11], [22], [23]. The later stage of the disease shows less common findings such as bronchiectasis, septal and pleural thickening, and subpleural involvement. These findings are expected in COVID-infected lungs and other viral infections or diseases. A patient with RT-PCR positive and the above-mentioned findings is considered for COVID-19 infected cases.

Therefore, this study aims to understand the CT image findings so that they can be incorporated into algorithms of ML and DL with a focus on (a) the role of chest HRCT in the diagnosis and prognosis of COVID-19, (b) significant radiological findings, and (c) its severity based on CT score. Further, in this research work, we characterized HRCT findings in 188 patients, presenting a retrospective study on the collected database from Marwari Hospitals, Guwahati, Assam, India. The findings of the database have been compiled in association with a team of radiologists from the Dr. Bhubaneswar Borooah Cancer Institute (BBCI). This study aims to appraise the usefulness of the spectrum of HRCT chest findings on COVID-19 cases and estimate the infection’s severity, to rule out the findings for further processing, and explore the possibility of incorporating different AI tools for the design of a predictive diagnostic framework for fast confirmatory screening of COVID-infected patients.

2. Chest CT and significance of radiological modalities

In this section, we highlight the importance of chest CT and the significance of radiological diagnostic aids in combating COVID-19 infections.

3. Materials and methods

The section details the data acquisition procedure, the criteria of the included and excluded data, imaging technique and interpretation, and the de-identification process.

4. Data statistics and interpretation results

CT scan images are extensively used as diagnostic aids for the treatment of COVID-19 infected patients. From the collected images, we have obtained the annotation and data classification based on the statistics for the mentioned retrospective study. This section highlights the statistics of the data, truly based on retrospective study to incorporate its applicability for the development of ML and DL algorithms, which are to be used as AI tools to create predictive models based on pure database patterns acquired from CT scanners. At the end of this section, a discussion based on the statistical analysis and a few of the limitations of the available data is brought to the fore.

The allocation based on age group and gender for both COVID and post-COVID is shown in Table 3 and Fig. 3. Among the COVID positive cases, 57.5% are male, and 42.5% are female patients. The distribution of patients is shown in the box plot of Fig. 3(A) separately for males and females. As per age groups, referring to the plot shown in Fig. 3(B) the highest number of sufferers were 30–60 years old, with 27.4% male and 23.3% female. The next group is 60–90 years old, with 24% male and 11% female. The minor sufferer group belongs to the age group of 30 years (05.5% male and 07.5% female). We also have one patient (male) belonging to the age group of 90. The CT scan of this case is shown in Fig. 4.

Similarly, in 42 post-COVID cases, 40.5% belonged to male patients and 59.5% to female patients. The box plot in Fig. 3(C) shows the corresponding distribution. As seen from the butterfly plot Fig. 3(D), the highest post-COVID cases belong to the age group of 60–90 years, with 19% male and 28% female, followed by 30–60 years, 16.7% male and 19% female, and the rest belonging to the age group of 30 years, with 4.8% male and 11.9% female.

Hence, from the available database, we may conclude that the highest number of COVID-infected patients belong to the age group of 30–60 years, with a high prevalence among male patients. The post-COVID cases show a high rate in the age group of 60–90 years old, with more female cases.

HRCT of the COVID symptomatic patients has been performed between the 4th to 14th days of infection for the 146 COVID cases. The 42 post-COVID cases underwent the scan due to breathing difficulties, low oxygen saturation (${\mathrm{SPO}}_2$), palpitation, weakness with body ache after testing RT-PCR negative post-infection. Out of 188 cases, it was found that 43.6% of them had bilateral lung involvement, while 48.9% had involvement in the right lung and 50.5% had involvement in the left lung; the details are given in Table 4 and Fig. 5.

As observed from Table 5, and Fig. 5(c) both upper and lower lobes (65%, with 50% to 25% male female ratio) of the left lung were the most commonly involved, followed by the right upper, middle, and lower lobes (54%, with 35% to 19% male female ratio), referring to Fig. 5(b). Next in the count is the left lower lobe (25%, $16:9$) and the right lower lobe (16%, $9:7$). Considering the regions as shown in Table 6, 48.9% of the cases had findings in both peripheral and basal regions of both the right and left lung, followed by 16% in the central region.

In this study, it is noticeable that 44 female cases (50.6%) had a normal HRCT scan with the absence of all the parameters taken into consideration, followed by 40 male cases (39.6%). The most common finding is the presence of multifocal, bilateral, peripheral GGO, which is significantly found in 40 male patients (39.6%) and 34 female patients (39.1%), followed by 7 (06.9%) male and 1 (01.1%) female with minimal opacities. The HRCT performed on the cases mentioned, with minimal to severe GGO was done approximately between the 4th to 12th days from being infected with the virus. The next common finding that was seen in 25 male (24.8%) and 16 female (18.4%) patients is the presence of consolidation. GGO with consolidation was present in 15 male (14.9%) and 6 female (06.9%) patients. A crazy-paving pattern was found in 6 male (05.9%) patients and 1 female (01.1%). GGO with the presence of a crazy-paving pattern was found only in 5 male patients (05.0%). A halo sign has been seen in 2 (02.0%) male and 3 (03.4%) female patients. None of the cases had all the seven parameter findings. Maximum cases had only one parameter finding (either the presence of GGO or consolidation or GGO with consolidation or GGO with crazy paving patterns), followed by a minimum of four to three-parameter findings [Table 7]

The CT score is done visually based on the severity of the disease [Table 8 and Fig. 6].

A total of 85 patients with COVID and post-COVID infection have no lung involvement. The highest CT severity is assigned to 1 patient with a CT value ranging between 20–25 with bilateral lung involvement, involving all the lobes [shown in Fig. 2(a)]. 40 patients had CT scores between 1 and 5, where the cases showed bilateral lung involvement in a few cases and involvement of either one lung in a few cases with the presence of minimal to moderate peripheral GGO, GGO with consolidation, and halo sign [refer Fig. 7 for an example]. 37 patients showed CT severity ranging between 6–10, showing patches of consolidation, consolidation with the presence of GGO, and a few crazy paving patterns in less than 10 patients, for example as shown in Fig. 8. 19 patients had CT severity between 11–15, involving both lungs with moderate GGO and consolidation, as shown in Fig. 9 as an example, followed by 6 patients whose CT severity ranged between 16–20 showing bilateral peripheral GGO with interstitial thickening, GGO’s with consolidation, and bilateral pleural effusion [refer Fig. 10 for an example]. Several chest HRCTs reported in our study with symptoms of chest pain, pounding heartbeat, shortness of breath and fatigue were found to have opacities parallel to the pleura [as an example refer Fig. 11], fibrosis along with bronchiectasis as shown in Fig. 12, interstitial thickening and subpleural band [Fig. 13 for reference]. The evidence of the prior information on HRCT confirms the presence of the virus in earlier months, along with some pre-existing diseases. However, a few cases with post-COVID symptoms did not show any findings on HRCT done 2 months prior to the infection. These few cases may have other pre-existing diseases due to which they might experience the above-stated post-COVID symptoms.

5. Discussions

As per the statistics of North-East India collected during the pandemic stage, thousands of new COVID-19 cases were registered on a daily basis. The gold standard for COVID identification, the RT-PCR test, is efficient, but it shows true negatives and therefore fails to be 100% accurate. Moreover, the case becomes more challenging for asymptomatic patients.

As a secondary measure, CT-scans are advised, but generally patients show less interest unless they are suffering from symptoms related to lung disorders. Also, the lung anomalies start at a later stage of COVID inception and when not treated in time, lead to total lung dysfunction and finally death. During the second wave of COVID, the post-COVID anomalies have increased the number of death counts.

As already mentioned, we have collected CT-scans of a total of 188 COVID positive patients. However, as far as the accuracy of the corresponding RT-PCR report is concerned, we have found two HRCT cases that show that they belong to COVID positive patients but were actually evaluated as RT-PCR negative. The CT scan of these two cases is shown in Fig. 14 which clearly signifies the presence of COVID. The database statistics showed more infections in the age group of 30–60, with a higher rate of male patients. This demography may vary depending on location and populations, but there is a probability match of the statistics with those recorded in other parts of India [7], [41]. Next, as per the HRCT findings on lung involvement, the pattern is similar in both genders with more effect on the upper–middle–lower lobes in the right and upper–lower in the left lung, respectively. As per the study findings, out of a total of 188 cases, 40 (39.6%) male and 44 (50.6%) female patients showed normal HRCT with zero findings with respect to GGO and consolidation. Thus, CT-scans cannot be used as a standalone measure for screening and RT-PCR will continue to be the gold standard for mass identification. But the point to be noted is that for the RT-PCR positive cases or post-COVID, CT-scans are mandatory. These studies were also reported in previous literature [11], [31], [41], [42], but the main objective of this work is to understand the HRCT findings useful to apply image processing tools and for automated AI algorithms to classify and grade COVID-19 severity from the samples.

Considering the versatility of the disease, the detection process based on radiological findings has many shortcomings despite its wide applications in diagnostic centers. Correspondingly, researchers related to the medical and computer fields use AI tools such as ML and DL models for analyzing these radiological findings. There were many systematic reviews which showed that ML and DL algorithms can be used for various disease detection and diagnosis processes [21], [43], [44], [45]. These algorithms have proven themselves to be efficient in the detection and classification of various diseases from different types of databases [46], [47]. Using different DL algorithms, studies have embarked on the identification and differential diagnosis of COVID-19. The DL-based architecture in the field of COVID-19 based on radiological findings helps to reduce false-positive and negative errors in the analysis of the disease, providing a fast and safe diagnosis system for patients. Fig. 15 provides an overview of a predictive model for the detection and diagnosis of COVID-19 using radiological modalities. Various studies reviewed that DL based detection for COVID diagnosis is classified into two categories: pre-trained models with deep transfer learning (DTL) and customized DL [20]. Pre-trained models use trained data in areas that have similarity with the content of the application. Pre-trained models with DTL provide the facility to accelerate convergence with network generalization other than the general training models, which requires more computation time. On the other hand, custom based DL techniques have specific applications and deal with feasible architecture development, which gives accurate performance due to their consistency. The custom networks use specific DL algorithms or hybridization of DL techniques with other fields of application like AI, such as data mining, ML, etc. [48]. Various popular architectures such as AlexNet, GoogleNet, MobileNetv2, Inception, InceptionV3, Xception, VGG-16, VGG-19, ResNet, ResNet18, ResNet50, ResNet101, Inception ResNetV2, DenseNet201, XceptionNet, CCSHNet, FGCNet are used in both of these models [49], [50], [51], [52], [53], [54], [55], [56], [57], [58], [59], [60]. Responses reported on these architectures by state-of-the-art methods are reproduced in Fig. 16 where pre-trained model comparison is based on its sensitivity of detection, whereas the customized methods are categorized as per its accuracy of detection [49], [50], [51], [52], [53], [54], [55], [56], [57], [58], [59], [60], [61], [62], [63], [64], [65], [66], [67], [68], [69], [70].

As per the literature, although efficient algorithms have been proposed in this area, there are certain limitations to these studies. As reported in our previous publication of COVID-19 review [71] there are two fold shortcomings, which includes, $\left(i\right)$ improper database and $\left(ii\right)$ limited database. Due to the noisy, incomplete, and unmarked data, issues like redundancy, sparsity, and non-availability of values are encountered during the training of classification modules. Again, the unavailability of accessible implies the application of limited local databases, which are very minimal compared to the actual patient data. This leads to underfit or overfit problems while applying DL. Apart from this, with the unavailability of standard databases, it becomes hard to calculate the proper efficiency. Hence, considering these factors and the modality under consideration, we would like to continue our work in the direction of automation of CT-scan grading using ML and DL tools, to increase the sensitivity and specificity of detection.

6. Conclusion

The HRCT findings of the chest play an essential role for radiologists in adequately diagnosing the COVID-19 infection. Mass screening by chest HRCT is a must, along with RT-PCR tests, to rule out better disease management. The presence of common imaging patterns of bilateral, peripheral, multifocal GGO, consolidation, and GGO with consolidation, in the initial phase of the infected cases, may help in the early detection of the disease. Asymmetrical and bilateral distribution of opacities of lungs is considered a benchmark for radiologists to detect COVID-19 pneumonia.

The current retrospective study analyzes the clinical characteristics and outcomes of 188 patients with COVID-19 and post-COVID infections. It identifies six HRCT findings based on age, gender, laterality of lung involvement, lobe involvement of lungs, region of coverage and lastly, CT severity score for proper assessment of the severity of the disease progression. The strength of this study is to assist researchers in exploring the radiological imaging patterns or findings of COVID-19 to develop knowledge-based systems using AI tools for detecting and diagnosing COVID-19. At the same time, it will be helpful for radiologists and medical institutes as an additional tool for the early detection of the disease. In the near future, we will analyze these findings for implementing AI tools to propose a novel automated algorithm for detecting and grading COVID-19.

Although HRCT outperforms the RT-PCR tests as a mass screening tool, for the time being, RT-PCR is regarded as the primary gold standard test as being cost-effective and able to detect subclinical cases due to the rising number of infections in the community level transmission.

CRediT authorship contribution statement

Upasana Bhattacharjya: Conceptualization, Data curation, Project administration, Formal analysis, Resources, Writing – original draft. Kandarpa Kumar Sarma: Project administration, Validation, Supervision, Review & editing. Jyoti Prakash Medhi: Conceptualization, Investigation, Project administration, Resources, Data curation, Supervision, Review & editing. Binoy Kumar Choudhury: Investigation, Data analysis, Review & editing, Supervision. Geetanjali Barman: Conceptualization, Data analysis, Resources, Writing – original draft, Review & editing.

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.

Data availability

The authors do not have permission to share data.