Many in this thread are saying there's no need to upload this information to the cloud because the transform can be performed locally on modern hardware, but that puts significant pressure on the business case. A model in the cloud is protected. Stored locally, that same model is potential shareware. If you're seeking to maximize the number of true positive diagnoses, then a viable business with low marginal costs is a better bet than an open, stagnant research model. Further, as others have noted, this information may already be stored widely in other clouds due to patient post-diagnosis search and browsing activity. Thus, adding another surface is only a linear change to the patient's privacy. Put simply, the incremental privacy change to the end user here may be worth it.