Intrinsic image dataset manifolds and their relationship to neural network generalization have recently begun to be studied for the common domain of natural images, but little such research has been attempted for radiological images, which is important due to the difference in relevant visual features for downstream tasks. Here we (1) compare the intrinsic manifold dimensionality (ID) of radiological and natural images and (2) investigate the relationship between ID and network generalization ability (GA). We find that radiological images generally have lower ID than natural images, yet a steeper correlation between GA and ID. However, to our knowledge, a formal model that explains this "domain shift" does not exist, which could be enabled by more rigorous mathematical analysis, leading the way towards more principled development of methods specifically tailored for medical image analysis.