That doesn't really matter too much from the implementing-ML point of view, you can just use it as a file store. DICOM files themselves are annoying too (especially if they bury stuff in private tags), as are HL7 (and EMR integrations) but .. that's mostly just work.

Agree the viewers lack flexibility but that's a lot more solvable than say the morass of EMR. If you are just looking at image interpretation visualizing things isn't so bad, if you had the models to visualize.