Radiology is the most liable field for medical negligence due to human error in the diagnosis of various diseases. According to John Hopkins University, an estimated 250,000 people in the United States alone die each year due to medical errors. While this is a disheartening figure, the technology we have available today makes it possible to change this statistic.
We at X-LEARN seek to harness the power of modern machine learning models to assist physicians in making accurate predictions for patient conditions including pneumonia, brain hemorrhaging and retinal abnormalities simply by uploading an X-ray image.
What it does
X-LEARN is a platform aimed at bringing insights on various patient aliments through CT Scan results and revolutionary machine learning. Through X-LEARN, medical professionals, researchers, students, and developers have the ability to upload an X-RAY image and obtain predictions on potential diseases and threats to the patient.
Users have the ability to register/login to our portal and upload a given X-ray image, upon which they will be given a prediction on whether the photo is of a brain, chest or retina. Furthermore, based on the initial body part classification, a prediction will be given along with the confidence score on whether the X-ray image indicates a given aliment with the patient (Pneumonia for chest scans, hemmoraging for brain scans, and Choroidal Neovascularization for retinal scans).
How we built it
X-LEARN is built through the power of Tensorflow and the Azure Custom Vision AI Tools which both delve into machine learning in the hopes of providing different data to analyze with. Through training an Azure Custom Vision AI model with datasets of patient X-ray images, our models are able to find patterns from different X-RAY Scans and Transfer Learning.
X-LEARN's core feature is its revolutionary machine learning model where it provides an accuracy rate of approximately 90% for test data across different ailments. Through this model, we expanded it such that we built a Serverless API through Firebase Functions and we also built a React App deployed on the Firebase Cloud. Furthermore, building the API of X-LEARN allows for other developers, researchers and the like to fully explore what X-LEARN is truly capable of.
Challenges we ran into
Initially, we decided to build our own machine learning model for image classification in a Jupyter Notebooks environment. However, upon training and testing our own model we discovered that its accuracy rate was not as high as we would have liked. Attending an Azure workshop provided us with the tools and knowledge necessary to instead utilize existing classification models built by Microsoft using the Azure Custom Vision AI platform. This allowed us to greatly improve our test accuracy score for brain hemorrhages and allowed us to delve deeper into exploring an exciting new AI platform.
Accomplishments that we're proud of
We were able to combine our unique talents, focusing on our individual areas of expertise to successfully deploy a project we initially believed was beyond our scope of capability. We were able to effectively utilize and build a visually appealing front end, and integrate backend functionality such as Firebase user authentication, and a Serverless Solution.
What we learned
The entire experience was an excellent learning opportunity. We learned:
- How to effectively utilize Azure Custom Vision AI to train powerful models
- How to train and test models through separation of image data
- How to integrate with Serverless Solutions within the Firebase Ecosystem
- How to implement effective UI/UX design within an application to improve user experience
What's next for X-LEARN
Additional features can be added in the future. For instance, a well-designed landing page would improve the marketability of our application. In addition, functionality can be created for doctors to add new patients and store uploaded images with test results for a given patient. Ultimately, in order to be marketed as a tool available in the real medical world, the accuracy of test results should be improved to the point of near perfection. However, we are proud of the initial working prototype we were able to create and the new ideas it has generated for us.