Inspiration

Currently covid-19 diagnosis is a difficult and slow process that is creating a problem in this terrible situation. Difficult because it requires expensive and slow diagnosis techniques (RT-PCR), that are not 100% accurate, or other quick and cheap test with much lower success rates.

Chest X-ray has been pointed out as an affordable and effective way to detect Covid-19 even in asymptomatic patients. But it can be also used as a tracking and prognosis, because it can identify the Covid-19 patient development, classifying the early, moderate or server patient stage. That allow to apply to the patient the precise medical protocols every moment plus assign and manage the hospital resource appropriately.

What it does

Covid19-RxDx it is approach by Supervised Learning suing machine learning / deep learning of artificial neural network algorithm with convolutional layers and applying transfer learning techniques of successful models with data augmentation and regularizing methods, in order to predict a binary classification whether a given chest x-ray image has Covid-19 (or not).

The Non-Covid-19 cases includes not only normal Chest X-ray but also any other pneumonia diseases, that has been proven it is well differentiated by the Covid-19 lung pattern disease.

Our model can perform multi-class classification providing 3 types of classes corresponding respectively to Covid-19, Normal and Pneumonia. In addition to that the model can classify the Covid-19 patient stage by early, moderate or severe prognosis, as mentioned above.

How we built it

The core of the tool is based on Google's tensorflow framework, together with Keras using python. All of them are open source therefore not licenses fees apply. In addition to general commitment from Google to guarantee maintenance for those open tools.

Challenges we ran into

The main problem of the platform is having access to a reliable database with x-ray patients with Covid19. Due to the biases encountered on all the public dataset images managed until now.

Accomplishments that we're proud of

At the moment, after the latest training of the algorithm yesterday, comprising 1262 training images, of which 631 are Covid-19 it excels with up to more than 99.5%% of accuracy, more than 99.5% sensitivity and more than 99.55 % of specificity, as it can be seen from our confusion matrix graph.

These high metrics are achieved after implementing fine-tuning procedures, on certain model architecture layers, that allow to train (or frozen) specific layers as a kind of final “nano neurosurgery” to our "brain" ensemble model, applied according to what more abstract or first basic feature extraction we want to be performed by our model.

That is to say, the higher performance published at the time being, all over the world ! So obviously we need to increase our datasets training, in order to get better validation increasing generalization and avoiding any biases presented on our publics and private sources of images applied.

We have end up with a web server prototype, to demo our Chest X-ray diagnosis concept, but due to strictly medical reasons that this tool have neither yet been validate nor accomplish to the all medical standards, including security, data protection and ethic impact, we might not be allowed to run the demo outside this hackathon test period. Please check it in the link below on "try it out" link or directly on http://www.covid19rxdx.es/

What we learned

Our multidisciplinary team contains from engineers to doctors, and we believe it is crucial to have on board all the individuals involved, starting with the radiologists.

This a critical issue because Radiologist are the key part of the validation and approval part. Also, we have learned that other medical instances such as emergency, etc. could be assisted with these AI tools, due to specific lack of knowledge on chest X-ray area but using frequently these images as standard protocols to get first patient tests.

What's next for Covid19-RxDx

Partner with other key partners. Already started Collaboration with Hospital and Health authorities to have access to more images and to radiologists to validate and improve the tool. As a matter of fact, we have starting a collaboration with Clinic Hospital of Granada (associated with Granada University) plus other entities such as Medical Association in Andalusian region and other private hospitals, where we are waiting to close a deal.

Recently our research team have been granted to access a huge clinical histories dataset, including many Covid-19 RX chest imaging, created and hosted by the private Hospital Group HM regarding "covid-data-save-lives" one of the best quality and biggest international dataset Covid-19, all over the world. (See https://www.hmhospitales.com/coronavirus/covid-data-save-lives/english-version)

In addition to that , our team is glad to announce to have closed an specific ad-hoc collaboration with CRIDA, see https://crida.es/webcrida/, a Spanish research reference center on ATM A.I.E development and innovation, to boost us our covid-19 dataset preprocessing capabilities.

Collaborate under healthy "coopetition" rules with other teams worldwide either as a complementary business model development or investigation approach to keep growing together for this enormous Covid-19 worldwide challenge!

Keep training the algorithm with more sources of chest x-ray images. Create a Proof of concept with CT-images. Train the model to predict different pneumonias moving the algorithm to multi-class classification.

Improve the web platform with new requirements from the users. Improve security of the platform and study more legal and ethical implications.

Built With

Share this project:

Updates

posted an update

We have already trained 631 chest X-ray images of Covid-19 vs another 631 Non-Covid-19 (associated to different pneumonias or normal images), using transfer learning techniques of a ML/DL algorithm and we got excellent performances (higher than 99% on accuracy, specificity and sensitivity)...we think we need to increase dataset training, particularly through the enrolment of hospital radiologist image signature to confirm that we are not having biases coming from the use of public datasets bank biased...this our main concern right now !

Log in or sign up for Devpost to join the conversation.