We made an disease detection machine learning model for medical images. It’s able to look at a medical scan (CT Scan, X-Ray, etc.) and highlight any regions that contain diseases or other anomalies.

It works by training a GAN on healthy images of lungs, and an encoder that’s trained to turn healthy images of lungs into latent space (compact number representation). The encoder is trained using cycle-consistency loss, which means that if the encoder can turn an image of a healthy lung into latent space, and the GAN can reconstruct that output into an image that resembles the original, then the encoder works well.

Once we’ve trained the GAN and encoder, we can test if a new image is healthy, and if not, where the disease is in the image.


What inspired you (or your team)?

It’s useful because existing machine learning models for disease detection can only find what they’re looking for. A breast detection model will never detect pneumonia if it’s in an image. Our project can detect any disease in a medical image, this will save doctor’s a ton of time, or catch their mistakes when they overlook something.

In a world where 10% of patient deaths are caused by misdiagnosis, we need a new way to detect disease that gets rid of human error. By using AI, we’re saving doctors time and preventing misdiagnosis.