This "commercial use" discussion is a bit nebulous and abstract until it is applied to a particular situation such as what Jorge describes.  Beyond the importance of the precise licensing terms and what previous court cases have interpreted, the enforcement of the license must be considered.  If nobody ever found out that the data were used to make money nothing would happen, and similarly if the licensor did not care enough to pursue legal action.  But what about the large institution where all of this is going on?  Their deep pockets make a good target for litigation so a big concern is to limit liability.  

Let's say a lab at the University College London acting as a CRO uses some data to create a pattern recognition algorithm that is used in a drug trial, and due to an error in the results of that algorithm, somebody dies.  Well-funded lawyers from the drug company will find somebody else to blame.  The error could have been caused by a bug in the algorithm, or by a mistake in the original data used to create the derivative work (perhaps a probabilistic atlas) that the algorithm used.  It would be expensive to find out.  The legal department's knickers get into a twist at even the possibility of a lawsuit.  I know this from my experience with Massachusetts General Hospital in licensing software for release as open source: the result was years of pain to get a restrictive and temporary license.

The question of, "is a derivative work allowed by the NC license condition?" is moot to a large organization: they would rather avoid it by either paying for the commercial license, or more likely, just say no you can't use that data for a drug trial because it is too risky.  They would have no problem with your using that data for scholarly purposes because the liability is less.

From my perspective here at Neuromorphometrics, I would be terrified to use those MICCAI Challenge scans for clinical purposes.  While each individual scan is as accurately labelled as possible, it is a small number of subjects in a limited age range.  I think those were mostly healthy young adults.  As we increase the age range of subjects in our labeled brain subscription, we're encountering new problems in both older and very young subjects, and we have only begun to look at scans of subjects with disorders, let alone any rare algorithm-breaking anomalies that we might find once we've quantified "normal" across all ages for all peoples.

On the other hand, my general terror is from not knowing what use/abuse might happen after we release the data.  If your algorithm really can extract some invariant from the data and it can be reliably and safely used to diagnose, or measure response to and guide treatment, that's great!  That's what we're all here for.  But the onus is on you to demonstrate that whatever is derived from the data is safe and effective.

Andy.