Vinidra diagnosis
The vision you deserve.
The problem Vinidra diagnosis solves
-
Tuberculosis (TB) is one of the major threats to human health worldwide, leading to millions of deaths every year. Tuberculosis is also the second leading cause of death (after HIV) in case of infectious disease. In the context of Nepal, NTP Anual report estimates total deaths of around 7.4 thousand in the year 2076. Although early diagnosis and treatment can greatly improve the chances of survival, it remains a major challenge, especially in developing countries like Nepal. Due to the immense abundance and affordability of computers, we can use computer-aided diagnosis to solve the problem.
-
Keeping this in mind, we tried to build a prototype that makes use of Tensorflow(Convolutional Neural Networks) and classifies X-ray images among the sick(potential TB), healthy and active Tuberculosis. We deployed our model in Django(web app) to emphasis accessibility for the users, especially in remote areas. We aim to make use of our prototype in such areas where there's the equipment, but no medical professionals to operate them effectively. It is a problem especially in Nepal because the authorities set up equipment in the hospitals through various fundings but medical professionals refuse to work in such remote areas. Our prototype may solve this problem because you don't have to be an expert in the field to operate an X-ray machine and feed the image in the web app. An x-ray machine operator(having no or less medical experience) can diagnose the disease with high accuracy.
-
Despite the remote areas, our prototype can even be used by professionals in big hospitals because even the best radiologist has an average accuracy of 66.7% (http://mmcheng.net/tb/) to diagnose tuberculosis through an X-ray image. We can proudly claim that our model has an accuracy of about 97% when checking with test data in TensorFlow. However, some percentages may decrease due to overheating and other challenges, but it will be significantly better than 66.7%.
Challenges we ran into
We couldn't apply the Tensorflow Object Detection API:
We tried really really hard to add an annotation feature in the TB prone areas in the image. We tried converting the dataset's XML files into TFrecords and make relevant pipelines for the detection, but due to some reason, it didn't work. After getting stuck for a long amount of time, we decided to ditch the annotation feature and add the colours in the images with the label's response. Overlap uploaded image with translucent green if healthy, orange if the model shows a sick lung, and red if the model shows TB. It might not be as good as the annotation feature but good enough for a 48-hour prototype.
Other minor issues:
Other issues weren't that much of a big deal. We had plenty of issues but picking up the errors, googling it, looking at the GitHub and StackOverflow, and making extensive use of Tensorflow, Keras and Django documentation solved all of it.
Lack of experience:
Since we're in the freshmen year of our college, we lacked expertise and good experience, especially in the Deep Learning domain. However, we were enthusiastic enough to do this project and it turned out to be okay. The time constraint was tight; however, our effort and enthusiasm helped us overcome it. I think that's the beauty of hackathons.
Technologies used