Inspiration

The global outbreak of Coronavirus (Covid-19) is hurting everyone around the globe. The hospitals are working at full capacity and most of them have very limited space to offer for the patients. If you show symptoms and suspect that you have the virus, you would go to the hospital. If clinical suspicion persists after the examination, a sample of nasopharyngeal exudate is taken to test the reverse-transcription polymerase chain reaction (RT-PCR). Getting the results of the PCR test may take several hours. While waiting the results, many hospitals around the world perform a chest X-ray. The chest X-ray is a discriminating element; if the clinical situation and the chest X-ray film are normal, patients can go home and wait for the results of the PCR test. But if the film shows pathological findings, patients should be admitted to the hospital for observation so that they do not go back home and spread the virus and so that they receive the necessary treatment at the right time. Unfortunately, there have been many misreading cases in which doctors send the patients that actually have coronavirus back home and keep the ones that do not have coronavirus at the hospital, causing a misallocation of scarce resources. These situations have contributed to the spreading of the virus and even caused unfortunate deaths. We believe we can make use of artificial intelligence as a pre-diagnosis tool. This could help using limited hospital sources efficiently and could contribute to contain the virus. This is why we came up with T-Covid: A Fast COVID-19 Diagnosis Tool.

What it does

In the age of information and technology, with the rise of artificial intelligence, AI started to be used in different types and places varying from health and medicine to education. T-Covid is a deep learning program that detects and classifies Covid-19 patients from healthy patients through Chest X-Rays with a testing accuracy of 92%. The model runs on a website that allows the users to upload CXR(Chest Radiographs) that need to be classified. After the image is uploaded, our program tells if the patient has the coronavirus or not within seconds. T-Covid not only detects Coronavirus from CXR, but it also contains crucial information about the disease and how to avoid it. It informs the users of the system about the current pandemic.

How we built it

In this study, since it requires vast resources to create a new dataset for this task, we seek for such a dataset available online to carry out our training with the chosen algorithm for only proof-of-concept. We first made literature research to find out the public datasets that we can make use of. For the research and development of our deep learning model, we collected Chest X-Ray images from 128 Covid-19 patients. The main task is to classify X-Ray images as "Healthy" and "Covid-19." First, to determine which approach fits our solution better, we analyzed the different state-of-the-art object detection and segmentation algorithms, such as R-CNN, Fast R- CNN, RetinaNet and Mask R-CNN with mostly used backbone networks. In our training strategy, we used IBM Watson's tool to come up with a model in such a limited time. We designed our deep learning model and virtual server images in a way that they can carry out deep learning tasks as fast as possible.

Challenges we ran into

One of the major challenges we encountered was the lack of data for CXR (Chest Radiographs) of Covid-19 Patients. Because Covid-19 is currently affecting people and because it is different from the other similar pandemics there was not a complete dataset with enough radiographs. We solved this issue by making our dataset. researching various international and local databases, reading articles and collecting as many datasets as we can from the internet. At first, we tried to implement a system that uses Computerized Chest Tomography of the patients. However, due to the lack of access to data on CT(Computerized Tomography)Scans we could not use CT implementation. Due to the lack of data, we trained artificial intelligence with 128 Covid-19 CXR and we tried to minimize the noise of the dataset however because we were not as skilled as medical experts on the radiography we could not make it zero. If the governments or health organizations provide these datasets publicly the T-Covid will perform more accurate and precise results with +95% accuracy.

Accomplishments that we are proud of

As we stated earlier, we had difficulties while finding a proper dataset. One of the achievements that we are proud of is the dataset we made to train this Artificial Intelligence model. The effort and time that we spend while collecting the images and removing noise from the dataset is one of the reasons why this dataset turned out to be accurate. One of the other accomplishments that we are proud of is the accuracy of the model. It was 92% percent accurate on the testing and training accuracy was even higher. But most importantly the greatest achievement we accomplished was the teamwork we made.

What we learned

Developing this system was a learning journey for our team. In the beginning, we did not know much about medical terms and phrases used in the articles about Covid-19. However, in the end, we learned a lot about Covid-19 disease and possible treatments and detection and classification techniques for a patient. We developed ourselves in various ways, and with the research that we made on the medical side of Artificial Intelligence and Covid-19, we gained pieces of knowledge over the Coronavirus from a medical perspective. We learned to become a team and work in an organized way to generate value through creative enterprise and hard work.

What's next for T-Covid: A Fast Covid-19 Diagnosis Tool

Artificial intelligence can impact many lives, and help us get through this crisis with less damage. However, the main problem is that there are not enough public datasets to carry out deep learning tasks that will fight with Covid-19. We need more samples from patients that have coronavirus so that we can make this program much more effective and accurate. As stated before due to the lack of access to data on CT Scans we could not use the CT implementation. One of the major improvements to this software is using Computerized Tomography Scans. Because in several types of research and articles it is stated that using CT(Computerized Tomography)Scan data improves the accuracy of the model we trained. With the support of governments or health organizations, T-Covid would be +95% accurate. After the improvements, T-Covid would be used as a pre-diagnosis tool that would help doctors to prioritize the diagnosis of patients with a high risk of coronavirus and start the necessary treatment earlier. We strive to create value together with people willing to fight with the global outbreak of coronavirus, as our collective efforts will ultimately make a real impact on the everyday lives of millions.

Built With

+ 19 more
Share this project:
×

Updates