With a collection of 300.000 X-ray images, deep learning networks can be trained to classify abnormalities effectively. The current algorithm can classify more than 15 different abnormalities. To improve the performance some prior knowledge can be implemented. Besides X-ray images other information is available:

Demographic Knowledge: With the help of other features like gender, age or smoking history, the network can increase the classification performance.

Lung Segmentation: Another network determines the area of the 2 lung lobes. The rest of the image can be sorted out to remove the non-important features before the classification network.

Spatial Knowledge: The location information of certain abnormalities is present e.g. nodule in upper left lung lobe.

This data collection is one of the biggest in the community. Therefore, experiments can be triggered to see the classification performance depending on the amount of training images. Those insights may also apply for other disease related projects.

Another task is to detect the abnormalities. Due to the unavailability of exact position, semi- and unsupervised learning algorithms are created. So-called attention maps can be predicted to heat up the region of interest.