pneumonia_detection

• 

  Home Page

• 

  Results

• 

  Transfer learning code. Dropout added to reduce over-fitting

• 

  Freezing the first 20 layers and training the last 20

• 

  Structure of the dataset used

• 

  You can test the model by uploading this image.

• 

  You can test the model by uploading this image.

Inspiration

The fact that we are in a global pandemic encouraged me to build something for the medical industry. That combined with my passion for computer vision gave birth to this project.

What it does:

The web app allows a user (usually a doctor) to upload a chest x-ray. And the Deep Learning model (Retrained MobileNet model) classifies the x ray into one of two categories, Normal and Pneumonia by comparing the probabilities. These probabilities are generated using the soft max activation function.

These predictions are then plotted for the user to visualise.

How I built it:

I used transfer learning to retrain the legendary MobileNet model to classify chest x-ray images I found on kaggle.

To do so I froze the first 20 layers and trained the last 20.

Once I built the model, I used my knowledge of flask and deployed the model as a web app using flask. I then deployed it Heroku.

Challenges I ran into:

The first time I trained the model, I ended up over-fitting. So I retrained it with fewer epochs to get more accurate results.

Also while deploying my app to Heroku, I kept running into errors because my build kept failing. It was because of my dependencies and their versions.

What I learnt:

I learnt to fix over fitting by reducing the number of epochs and adjusting dropout.

I also learnt to debug issues while deploying my flask app to Heroku.

Built With

• flask
• html
• python