Context

When I started playing around with deep learning in radiology, the first barrier I faced was obtaining a dataset. So, I just downloaded some public images from google images.

Content

This dataset contains 100 normal head CT slices and 100 other with hemorrhage. No distinction between kinds of hemorrhage. Labels are on a CSV file. Each slice comes from a different person. The main idea of such a small dataset is to develop ways to predict imaging findings even in a context of little data. In this notebook, I present a simple data augmentation capable of achieving 90% accuracy in the test set.

Acknowledgements

Thanks for the people who made their images available on google.

Inspiration

Help push the frontiers of Artificial Intelligence in Medical Imaging.