Latin America is facing a great challenge due to current situation with COVID-19. At our country, for instance, hospitals are slowly running out of beds, people is getting infected at an increasing rate and governmental politics have shown this matter is not easily solvable. That's why we are trying to contribute by setting the first steps on a model that can help medical staff to diagnose COVID, as it can be confused with other diseases as pneumonia.
What it does
The project itself its divided into two parts, the Jupyter notebook worked in google collaborate contains both algorithms. The first tries to classify images (slower and much more computationally demanding) meanwhile the second contains the code of the graphic interface we used. It can classify a torax plate and tell whether it is an example of a covid pneumonia, a non-covid pneumonia or a healthy lung.
How We built it
We used several libraries in the process. We used a Keras model based on CNN algorithm in the graphical interface, despite its high computational cost, and, after setting the buttons that allow user to select any image from its computer (with the functional requirement that images need to be in both .jpg and .gif format), we achieved the presented the final functional interface.
Challenges we ran into
First of all to import the data. It was challenging to import thousands of images into arrays so that they can be analysed. We tried to use an algorithm that could set deep features but, as this wasn't possible (mainly due to time), we implemented the CNN and used the keras-utils library instead. We thought the challenge was coding the machine learning algorithm but in the end that wasn't a problem, but rather achieve a functional model with good accuracy. Also RAM was hard to control.
Accomplishments that I'm proud of
The CNN model works nicely, even if due to RAM its hard to use it in our computers. We think it can help other projects on their development.
What We learned
In this project we learned to develop and strengthen our programming skills. Additionally, we managed to carry out a project that solves a problem that has a great incidence in Colombia. We also learned how to build these type of prediction models, such as neural networks.
What's next for Maching Anding
The idea would be to implement the same algorithm with more data so that it classifies more illness and more type of medical images. It depends on whether we find datasets or people with willing to cooperate with this project.