Waterkant C10 Challenge Team fhp

Inspiration

In the current situation of COVID-19 the assessment of radiograph images of lungs can help to optimize treatment for patients. Artificial intelligence may be able to improve diagnosis of COVID-19 patients based on analysis of chest radiograph images.

What it does

The final neural network takes a radiograph image of a lung as an input and predicts whether there are signs of pneumonia of not.

How I built it

In the first few iterations we tried to assemble a neural network with some convolutional layers for image recognition. Later on we switched to use a pretrained model based on imagenet. We omitted the last few top layers of the existing model and added our own layers, which then were trained on the trainings set of radiograph images.

Challenges I ran into

One major challenges were the inexperience we had when working with neural networks and transfer learning in particular. Another great challenge was to make the right adjustments to the pretrained model in order to improve its performance. Additionally as in all AI challenges time is a very limited factor. Especially when the training of a specific model takes up to a few hours.

Accomplishments that I'm proud of

We managed to get up to 70% accuracy on the prediction of pneumonia, although we only had a few epochs to train the network.

What I learned

We learned some basics of applying neural networks to image classification problems. Especially constructing a neural network from ground up and using pretrained models and adjusting them to fit our problem.