Yes, and that's possible now with many different CNNs. The limiting factor is the training/validation/test data in your subject.

For example you couldn't implement Mask R-CNN with the COCO dataset as implemented here and get inference on your radiology problem set.