Diagnosis of COVID-19 using CXRs and CT scans

Abnormality Detection in Musculoskeletal

-Rohan Kalluraya

10 June, 2024

Coronaviruses are a large family of RNA viruses that are known to cause respiratory tract illnesses like the common cold. Most coronaviruses affect animals and can be transmitted between animals and humans. COVID-19, symptomized by fever, cough, and shortness of breath, is the latest addition to the group. The real-time reverse transcription-polymerase chain reaction (RT-PCR) test is the most commonly used reference test for the diagnosis of COVID-19.


This blog article describes the work that is published in detail at:


Rohit Lokwani, Ashrika Gaikwad, Viraj Kulkarni, Aniruddha Pant, Amit Kharat. "Automated Detection of COVID-19 from CT Scans Using Convolutional Neural Networks"




The RT-PCR test comes with its shortcomings, which include the complex process of specimen collection, the long time required for the analysis, variability in the accuracy of the test, and a shortage of testing kits. RT-PCR tests can be supplemented with computer-based detection methods like machine learning that extract features from an image and give diagnostic outcomes. The most common and economical of these is chest radiography. Radiologists use chest X-ray images (CXRs) to detect pathologies like pneumonia, tuberculosis, and early lung cancer.


But, detecting COVID-19 using CXRs is challenging due to the less evident visual features in CXRs caused by the overlapping of ribs and soft tissues and low contrast. Using chest CT scans, with their high accuracy and speed, resolves these shortcomings. It is essential to detect the presence of COVID-19 at an early stage to curb human-to-human transmission. Here, CT assists in the detection of minor infections. Our proposed method uses neural networks to identify the pixels showing COVID-19 affected regions in a CT scan and triage the patient as either high suspicion of COVID-19 or negative.


We trained a convolutional neural network to segment out the region of COVID infection and, in turn, separate COVID-19 cases from non-COVID cases. We used transfer learning to initialize this model with the weights from our COVID-19 vs non-COVID X-ray model, which helped us in improving the accuracy and reducing the time for optimization.



(Left) Original image and (right) predicted mask by 2D model.

(Left) Original image and (right) predicted mask by 2D model.


We evaluated COVID-positive, COVID-negative, consolidation, and healthy CT scans from multiple geographies globally including India. The ground truth in these images was decided based on their RT-PCR test results. The CT slices were annotated, classified, and marked positive by a group of trained expert radiologists. We observed a sensitivity of 96.3% with a specificity of 93.6% while evaluating the model. The dice coefficient on positive samples was 56.1%. We tested our model on datasets from sources across three different countries (China, Italy, and India) and found that the performance of the model remained consistent.


The diagnosis of COVID-19 using CXRs and CT scans has gained significance since the ubiquitous spread of this disease. Chest CT scans usually tend to show the region of infection more clearly than CXRs. In conclusion, chest CT has proved to have a higher sensitivity for COVID-19 diagnosis than CXRs. Our analysis makes a strong case for using chest CT for COVID-19 screening and evaluation, especially in epidemic situations, where community spread calls for speedy diagnosis. Chest CT can also assist in ascertaining the overall extent of infection and can have a potential role in disease prognostication.

Share this

Related Posts