Inspiration

When I was 12 years old my grandma got misdiagnosed. This issue complicated her pneumonia and almost took her life. I entered this hackaton knowing that I wanted to address this issues. Me and my team are believe in creating projects with a positive social impact. So we combined our knowledge of computer science and AI with our passion for healthcare to create S.M.A.I.R. VISION.

What does S.M.A.I.R. VISION do

S.M.A..I.R. VISION is an artificial intelligence algorithm that aims to detect anomalies in medial images and aid medics across the world in their diagnosis to make it faster and more accurate. It works as a tool for the doctor that allows him or her to take a more hands on approach with the patient while the computer analyzes the medical image. It also reduces the chances of medical negligence as saves the doctor time since in this current stage of the pandemic, a lot of medical services are provided online. This tool can be applied into telehealth and provide doctors with better tools for an excellent diagnosis.

How we built it

S.M.A.I.R. VISION is built through a convolutional neural network. The algorithm used has 3 hidden layers and was trained with 47000 medical images. This dataset is divided into train set and test set. The train set is constituted of 75% of the original dataset and the test set constitutes 25% of the original dataset. The algorithm has an 87% accuracy rate and a mean squared error of of 0.78 with a confidence interval of 5% in accuracy and a confidence interval of 0.12 in the mean squared error. The dataset ran through 5000 epochs in order to ensure that it was properly trained we also managed to get a ratio of 3:1 false positives against false negatives through experimentation and parameter tuning.

Challenges we ran into

Our first challenge was gathering the image, but upon contacting the Cancer Imaging Archive we managed to get 50000 images of which around 47000 were usable. The next challenge we faced was training our algorithm. None of our laptops could handle such training in the desired amount of time so we gathered some coupons from amazon web services and we train out neural network in the cloud and then imported the trained weight into our python code as html files. The last challenge was moving the parameters to minimize the number of false negatives. Since this is a medical product false positives are way more desirable than false negatives and we managed to get a ratio of 3:1 in regards to false positives and false negatives. We sacrificed some accuracy for this, but it is the correct thing to do.

Accomplishments that we're proud of

We are proud of achieving such a high accuracy and a low mean squared error in a short period of time. This is a genuine advancement for us and we had no idea we could get this far. We are also extremely proud of creating a product with high social impact since we questioned some local clinics and they 3 out of 4 mentioned that this product could make their life easier.

What we learned

We gained extensive knowledge on cloud computing and now understand more about the process of training a neural network and using servers from all around the world to our advantage. We also gained a deeper understanding of the function inside each neuron and the parameters since we had to do some intensive tuning to achieve the desired results. Overall we also learned that failures shape the path to success and through extensive experimentation, perseverance and coffee we managed to get the desired results.

What's next for S.M.A.I.R. Vision

We are really exited for what the future holds a head. This hackaton has filled us all with a renewed interest for artificial intelligence in healthcare. We are going to associate with the Santa Fe public health office to start during trials with our algorithm. We are also going to provide the project as a free open source initiative so that people in our community and all over the world can use it and even improve upon it to create more amazing projects. We believe there is a bright future ahead filled with opportunities for the application of our project and we will start by changing the name into something more catchy and easier to type.

Built With

Share this project:

Updates