Keywords artificial intelligence - computed tomography - COVID-19 - flu
Introduction
Coronavirus disease-2019 (COVID-19) was declared a pandemic worldwide in 2020. It
is not only contagious and causes acute respiratory distress but also multiple organ
failure in severe cases.[1 ]
[2 ]
[3 ]
[4 ] Apart from its prevalence and being contagious, its diagnosis poses a great challenge.
Reverse transcriptase-polymerase chain reaction (RT-PCR), which is considered the
gold standard for diagnosis of COVID-19, is mired with issues of low sensitivity of
66% with a high false-negative rate of 33 to 66% along with a false-positive rate
of 12 to 16%, apart from the fact that the test is time consuming.[5 ] The role of imaging in the diagnosis of COVID-19 is also under debate. Computed
tomography (CT) has a high sensitivity of 97%[6 ] and is a valuable imaging tool for the clinical diagnosis of early-stage COVID-19,
especially when there are insufficient RT-PCR tests as was used in Wuhan in the early
pandemic period.[7 ]
[8 ] CT has been shown to detect lesions in the early stage of the disease even when
RT-PCR may be negative;[9 ] however, it has a low specificity of 25 to 35%, and it may be difficult to distinguish
cases of COVID-19 from other viral cases of pneumonia and similar look-alike lung
pathologies.[5 ] American College of Radiology (ACR) Guidelines, however, do not relegate it as the
frontline test to diagnose COVID-19 primarily due to risk of contamination of radiology
departments and risk of exposure to health workers.[10 ] Hence, there is a need to add more tools to the use of CT to make it more specific
for the detection of COVID-19. Artificial intelligence (AI) is an attractive tool
to achieve this need. It has been used in many facilities to make the diagnosis of
COVID-19 more specific and has been shown to have an accuracy of 90%.[11 ] The use of AI not only speeds up the diagnosis but also has the potential to calculate
disease severity by performing accurate lung segmentations. As most of the countries
around the world are dealing with this pandemic by setting up COVID-19-dedicated hospitals,
at the same time still a majority of health care providers constitute non-COVID-19
health care facilities that continue to face the dilemma as to how to triage patients
coming with flu-like symptoms while facing constrained resources including manpower,
personal protective equipment kits, lack of free testing, and above all the fear of
contracting the contagion while treating such patients. Chest radiographs still form
the frontline screening tools in non-COVID-19 hospitals even though they have poor
sensitivity in early detection of COVID-19.[12 ] AI tools have now been developed to make chest radiographs efficient in the detection
of early lung changes so that rapid treatment and triage can be done in such patients.[13 ]
[14 ]
This study was designed to test algorithms using AI along with chest radiograph and
CT imaging as frontline tests to triage patients coming to non-COVID-19 hospitals
with flu-like symptoms ([Fig. 1 ]).
Fig. 1 Proposed imaging algorithm for flu-like presenting patients. CAD4COVID, computer-aided
detection for coronavirus disease-2019; CXR, chest X-ray; HRCT, high-resolution computed
tomography; ICU, intensive care unit; RT-PCR, reverse transcription polymerase chain
reaction.
Materials and Methods
The study comprised 3,235 consecutive patients with flu-like symptoms of fever, sore
throat, and cough with constitutional symptoms or with respiratory discomfort who
were referred to our institute from various primary and secondary health care facilities
for imaging and diagnosis in a period of 240 days, that is, from March 2020 to November
2020. All the patients underwent flu diagnostic evaluation as per [Fig. 1 ]. The chest radiographs were done on portable X-ray system, which was a makeshift
arrangement near the entrance of the institute. History and informed consent were
obtained from all the patients along with a history of any contact or any recent travel
outside the district. Body temperature was recorded of all the patients with a thermal
scanner. Under all protective gear, the chest radiograph was obtained in posteroanterior
or anteroposterior views and images digitally processed. These were then transferred
to computer-aided detection for COVID-19 (CAD4COVID) AI software (Delft Imaging Hwetogenbosch,
the Netherlands) and CAD4COVID threshold determined on heat texture image map. The
patients having a threshold score of <50 were classified as group A and were put on
symptomatic treatment and follow-up. Group B patients were those with a threshold
score of >50 and were labeled as positive for an underlying lung abnormality and underwent
plain high-resolution computed tomography (HRCT) examination on 128 slice scanner
(Siemens Go Top, Erlangen Germany AG). The images were transferred to COVID-19 AI
software (Quibim, Valencia, Spain; Thirona BV Nijmegen, the Netherlands). A threshold
of 50 was based on T score = 50 based on the area under the curve with threshold 50
having a sensitivity of 91% and specificity of 87%. The patients with COVID-19 similarity
threshold of >0.85 were labeled as positive for COVID-19 disease. The CT severity
scores were determined in these patients based on the percentage of lobes of lung
involved with 0 to 25% as score 1, 25 to 50% score 2, 50 to 75% score 3, and 75 to
100% score 4 for each lobe of the lung and totally added to form a total severity
score with the maximum score being 20. The percentage of lung involved by COVID-19
was also determined. All patients with a score of less than 9 were transferred to
COVID-19 isolation wards and those with scores more than 9 to COVID-19 intensive care
units for further management. The patients with COVID-19 thresholds of less than 0.85
were labeled as non-COVID-19 cases of pneumonia and transferred to non-COVID-19 medical
wards. The follow-ups were done as per [flowchart 1 ].
Results
The patient demographics are enlisted in [Table 1 ]. The mean age of patients was 39.5 and 51.2 years in groups A and B with 80 and
85% being males, respectively. The commonest symptoms seen were fever (79%, 86%) and
cough (42.8%, 63%) in group A and group B patients, respectively, followed by constitutional
symptoms, respiratory discomfort, and malaise with loss of appetite. Comorbidities
were seen in 60% of the patients; 89% of group A and 75% of group B patients had normal
total leucocyte counts, while lymphopenia was seen in 38% of group B patients. A total
of 91% of group A patients and 61% of group B patients had normal SPO2 levels. Group A patients formed the majority of patients, that is, 2,209/3,235 ([Fig. 2 ]) with a mean CAD4COVID score of 35 ([Fig. 3 ]). There were 22 false-positives due to artifacts of overlying scapulae and 28 false-negative
patients who had focal ground-glass opacities with a threshold less than 50. The pattern
of distribution of lesions in group A patients is enlisted in [Table 2 ]. Group B patients comprised 1,026 patients with a mean CAD4COVID score of 72 and
underwent plain HRCT examination with COVID-19 AI analysis of images ([Fig. 4A ] and [B ]). The patterns of distribution of lesions on plain radiographs and HRCT are enlisted
in [Table 3 ] and [4 ]. A total of 825 (25.5%) patients of COVID-19 were detected by COVID-19 AI analysis
using a threshold cut-off of 0.85 ([Fig. 5A ] and [B ]). The distribution of various cut-off scores is shown in [Fig. 6 ]. A total of 201 patients had non-COVID-19 diseases, as tabulated in [Table 5 ] and [Fig. 7A ] and [B ]). RT-PCR was done in 530 patients of whom 380 were positive and 150 false-negatives.
A total of 35 patients were detected positive on repeat RT-PCR after 2 days. In 12
patients, RT-PCR could not be done as they died. A total of 554 patients had CT severity
scores of less than 9 ([Fig. 8 ]) while 271 patients had scored more than 9 ([Fig. 9A ] and [B ]). Correlation of these scores was done with a clinical respiratory status of patients
([Table 6 ]). The patients in both groups were triaged as per the depicted [Fig. 2 ]. The follow-up data were collected in 2,187 patients of group A, all of whom showed
complete recovery while the remaining 22 patients were lost to follow-up. In group
B, 11 patients of COVID-19 died on the first day of admission while hospital mortality
was seen in 32 more patients of COVID-19 and 13 in non-COVID-19 category. CAD4COVID
analysis on plain radiographs at a threshold of 50 achieved sensitivity and specificity
of 97.9% and 99%, respectively, in the present study to detect any significant chest
pathology ([Table 7 ]). The overall prevalence of chest infection detected by the use of CAD4COVID analysis
in patients with flu-like presentation in the present study during the pandemic was
30.5%, with an incidence of COVID-19 being 25.5% in the present time period of 8 months.
This prevalence, however, varied in each month depending upon the progression of the
pandemic ([Fig. 10 ]).
Table 1
The patient demographics of both groups
Fig. 2 Algorithm with patient numbers in the study. CAD4COVID, computer-aided detection
for coronavirus disease-2019; CXR, chest X-ray; HRCT, high-resolution computed tomography;
ICU, intensive care unit; RT-PCR, reverse transcription polymerase chain reaction.
Fig. 3 Plain radiograph chest and CAD4COVID analysis image in group A patient showing threshold
score of 38. CAD4COVID, computer-aided detection for coronavirus disease-2019.
Fig. 4 (A ) Plain radiograph chest and CAD4COVID analysis image in group B patient with a score
of 60. (B ) HRCT axial image and COVID-19 analysis showing COVID-19 similarity of 0.99 in COVID-19
patient. CAD4COVID, computer-aided detection for coronavirus disease-2019.
Fig. 5 (A ) Group B patient with a plain chest radiograph and AI map with CAD4COVID score of
78. (B ) Same patient showing HRCT coronal image with color-coding of COVID-19 lesions with
COVID-19 similarity of 0.99 and a severity score of 8. AI, artificial intelligence;
CAD4COVID, computer-aided detection for coronavirus disease-2019; COVID, coronavirus
disease; HRCT, high-resolution computed tomography.
Fig. 6 Pie chart showing the distribution of coronavirus disease-2019 (COVID) similarity
values in group B patients.
Fig. 7 (A ) Plain radiograph chest and CAD4COVID score of 61 in a Group B non-COVID-19 patient
of H1NI pneumonia. (B ) HRCT chest with COVID-19 analysis showing focal GGO in the left lung with COVID-19
similarity of 0, ruling out COVID-19 as a cause. CAD4COVID, computer-aided detection
for coronavirus disease-2019; COVID-19, coronavirus disease-2019; GGO, ground-glass
opacities; HRCT, high-resolution computed tomography.
Fig. 8 Group B non-COVID-19 patient with pulmonary edema and pleural effusions with ground-glass
opacities with COVID-19 similarity of 0.49 on AI with a severity score of 9 triaged
to non-COVID-19 ICU. AI, artificial intelligence; COVID-19, coronavirus disease-2019;
ICU, intensive care unit.
Fig. 9 (A ) Group B COVID-19 patient showing CAD4COVID score of 89. (B ) Same patient showing HRCT AI analysis with COVID-19 similarity of 1.0 with a CT
severity score of 18 categorized as severe disease and triaged to COVID-19 ICU. AI,
artificial intelligence; CAD4COVID, computer-aided detection for coronavirus disease-2019;
COVID-19, coronavirus disease-2019; CT, computed tomography; HRCT, high-resolution
computed tomography; ICU, intensive care unit.
Fig. 10 Bar chart showing the monthly distribution pattern of coronavirus disease-2019 (COVID-19)
and non-COVID-19 infections.
Table 2
The pattern of lesions seen in group A on plain radiographs
Table 3
The pattern of lesions seen in group B on plain radiographs with CAD4COVID score.
CAD4COVID, computer-aided detection for coronavirus disease 2019
Table 4
The pattern of lesions seen in group B on high-resolution computed tomography with
severity scores
Table 5
The disease distribution in group B patients
Table 6
The correlation of computed tomography severity scores with severity scores
Table 7
The sensitivity and specificity analyses of CAD4COVID analysis on plain radiographs.
CAD4COVID, computer-aided detection for coronavirus disease 2019
Discussion
During an outbreak of a highly infectious disease like COVID-19 with a person-to-person
transmission, health care personnel have increased workload and work stress due to
limited experience and available tools to triage these patients. Although Pan et al[7 ] have shown that CT scan can show findings in the early and asymptomatic phase, another
study by Chung et al[15 ] has shown that CT may be normal in the early stages of the disease and can limit
radiologists' ability to screen asymptomatic patients. However, in the symptomatic
phase, that is, from day 5 to day 10 following infection, radiology is fast and very
accurate in forming the diagnosis of COVID-19 with 97% sensitivity of CT. Majority
of these patients with mild to severe flu-like symptoms present first to non-COVID-19
hospitals for which diagnosis and triaging these patients is a big challenge. This
study comprised a cohort of 3,235 patients evaluated with the use of AI along with
plain chest radiographs and HRCT as frontline tools to diagnose and triage these patients
into no chest infection and those with infections of COVID-19 and non-COVID-19 types.
The algorithm used in the study was designed to combine the benefits of lower cost,
quicker results, high sensitivity, and improved accuracy of both CT and plain chest
radiograph using AI. RT-PCR, which has problems of limited availability, longer test
times, reduced sensitivity, high false-negatives of 33 to 66% resulting in high nosocomial
infections and social distress, and lastly issues of repeated testing, was kept as
second-line test for confirmation of COVID-19 diagnosis. The present study was started
in the early phase of the pandemic when 87% of patients presenting with flu-like symptoms
to general health care non-COVID-19 facilities still had non-COVID-19 type of etiology.
These months being the onset of spring season had a predominance of seasonal flu due
to changes in weather; however, the study showed that as the pandemic progressed not
only there was an increased number of patients, but also 25% of these were those of
COVID-19 disease as is shown in [Fig. 10 ]. The study showed that combined use of plain radiographs of the chest with AI-based
analysis like CAD4COVID (at a cut-off threshold score of 50) with HRCT and AI analysis
could detect COVID-19 disease with high sensitivity and specificity of 97.9% and 99%,
respectively, in these symptomatic patients. The use of plain radiographs as a first-line
tool in non-COVID-19 facilities in such patients not only saves cost but also reduces
unwanted use of CT as a blind screening tool for all flu patients. The present study
shows that this algorithm is faster and more accurate than the conventional path of
using RT-PCR as a frontline tool to screen all flu patients. Only a few studies have
been done so far using AI with COVID-19 detection by imaging.[11 ]
[16 ] The study also shows that estimation of CT severity scores had a good correlation
with respiratory status of patients, especially with correlation with PAO2 /FiO2 values, and can help to decide the intensive care strategy in these cases. We used
a cut-off severity score of more than 9 to categorize as a severe infection. Yang
et al[17 ] also showed in their study that CT severity scores had sensitivity and specificity
of 83% and 94%, respectively, to detect severe COVID-19 disease. Follow-up CT if required
in these patients having a baseline pretreatment severity score can also be of help
to monitor the progress of the disease. Based on the initial experience of the use
of this new tool of AI combined with plain radiograph and CT, this study shows that
imaging can play an important role in the diagnostic workup of patients presenting
with flu-like symptoms. It can also predict the severity of the disease, which can
influence the prognosis and follow-up of these patients. Despite the pandemic, there
are a large number of patients who are non-COVID-19 and need diagnostic services in
non-COVID-19 hospitals. Chadha et al[18 ] have shown that the burden of seasonal influenza is not recognized in India, and
there is no reliable national data registry. The absence of data does not mean an
absence of seasonal influenza even during times of COVID-19 pandemic. It is known
that influenza disease exists in India, the types and subtypes of strains circulating
in the country, and the seasonality of annual outbreaks are also known. Hence, there
is a likelihood of concurrent outbreaks of COVID-19 and non-COVID-19 disease in the
country depending upon season and geographic location. Ongoing pandemic has currently
raised fear and anxiety in patients with flu-like symptoms, and we have seen an increasing
trend by patients to report to non-COVID-19 hospitals that were not seen in prepandemic
times, thus posing a diagnostic challenge. Blind use of RT-PCR screening to evaluate
all patients with flu-like presentation would be a waste of resources, more time consuming,
and add more confusion in patient management as is being observed in many patients
due to high false-negative and positives of RT-PCR. Using this approach alone would
also fail to triage other chest infections. At a 25.5% prevalence of COVID-19 disease,
this study has shown that with sensitivity and specificity of 97% and 99%, respectively,
and using AI, majority of the cases would be correctly flagged as abnormal, and 0.9%
(7/825) positive cases would be missed by this algorithm, these results being better
than RT-PCR done alone. Similarly, with a specificity of 99%, the likelihood of false-positive
is also negligible. Radiology has been the backbone of diagnostic services, and it
would not be prudent if these diagnostic examinations are not utilized as frontline
modalities provided they are used with proper protective measures. If correctly interpreted,
the guidelines issued by ACR[10 ] should be applied to only COVID-19 hospitals that treat established RT-PCR positive
patients where imaging can be supplanted with portable chest radiographs whenever
required.
To conclude, this study demonstrates that the use of AI-based analysis using chest
radiographs followed by HRCT with AI analysis is an accurate, cost-effective, and
quicker way to evaluate patients with flu-like presentation and helps to triage them
for early and proper management. This algorithm of evaluation for asymptomatic population
may be even more valuable and economical in countries with limited RT-PCR testing
resources.
