The lungs segmentation via UNet model were performed first to take a closer look on data distribution.

Resulted ROI were used then as describtors for clusterisation:

This two resulted clusters actually caught the difference of images came from different sources. Below are random samples from each of both clusters:

To overcame the domain difference an augmentation technique based on work of Philipsen et al has been employd enorm.py:

Finally RetinaNet model was used along with focal loss: