Brain Tumor Detection

💡 Inspiration

We wanted to reduce the time that a doctor would require to identify Tumor in Brain MRI Scans so we built a tool that could assist doctors in this task.

💻 What it does

Brain Tumor Detection first classifies a Brain MRI image as Tumor/Non-Tumor using a simple Convolutional Neural Network.
If the tumor is found, it further tries to localize it using a UNet CNN model.
Both the models are built from scratch.

⚙️How we built it

• ML: Python, TensorFlow
• Frontend: HTML, CSS, JS
• Backend: Django
• Database: CockroachDB
• Authentication: Auth0
• Sending results: Twilio

About the Model

Classifier Model-

• Dataset had 3929 total Brain MRI images
• 2556 images belonged to the category of No Tumor
• 1373 images belonged to the category of Tumor
• The model was designed using 6 Convolutional Layers, 1 Dense Layer and 1 output layer (with 2 neurons)
• Accuracy Archived - 96.27%

Localization Model-

• Dataset had 1373 total Brain MRI images
• 1167 images were used to train the model
• 103 images were used for Validation and Testing Respectively
• The model has a typical UNet Architecture with
  • 4 Downsampling Convolutional Blocks
  • A Bottle-neck Convolutional Block (without Maxpooling)
  • 4 Upsampling Convolutional Blocks with skip connections attached
  • 1 separate Conv2D layer
  • Conv2D Output layer
• Accuracy Archived - 97%

Use of Twilio

• For sending the result from the model to mobile number.

Use of CockroachDB

• We have used CockroachDB as a primary database because it is an easy-to-use, open-source and indestructible SQL database.

🔑 Auth0

• We have used Auth0 for secure user authentication

🧠 Challenges we ran into

This was the first time we were working with the image segmentation task and the problem with UNet is that since you have to manage those "skip connections" and add them later on, the dimensions of image matrices on both sides should be equal. We had to cover a lot of math to finally get those dimensions right and it took most of our time.

🏅 Accomplishments that we're proud of

We are elated that we were able to cover new grounds and learn something new in a matter of just a few hours and completing our project always gives us great pleasure.
Achieving around 97% accuracy on both the tasks was a cherry on the top.

📖 What we learned

We learned all about the UNet Architecture of Neural Networks that is primarily used for pixel-level image segmentation, collaboration, and teamwork.

🚀 What's next for Brain Tumor Detection

Improving the accuracy of both the models.

Built With

• auth0
• css3
• django
• html
• javascript
• python
• tensorflow