A dual-level evaluation framework for anatomical data augmentation: A comparative study on L2 vertebra semantic segmentation

Data augmentation is widely used to mitigate the limited availability of annotated datasets in medical image analysis. However, its evaluation is still largely based on downstream task performance, which does not necessarily reflect the anatomical validity or representativeness of synthetic samples. This study proposes a dual-level evaluation framework for anatomical data augmentation that explicitly separates geometric–statistical similarity from morphometric plausibility. The first level compares real and synthetic shapes in point distribution model space using complementary distributional and geometric metrics, instantiated here through Maximum Mean Discrepancy and Chamfer distance. The second level evaluates each synthetic instance in an interpretable multivariate morphometric space using the Mahalanobis distance with respect to a real anatomical reference distribution. The framework was applied to L2 vertebra semantic segmentation and used to compare three representative augmentation strategies: KNN+SMOTE interpolation, variational autoencoder-based generation, and statistical shape model-based augmentation. Results show that augmentation methods exhibit distinct and partially decoupled behaviors across the two evaluation levels. KNN+SMOTE achieved high geometric–statistical similarity to the development dataset but generated samples concentrated in the central region of the morphometric space. The variational autoencoder produced anatomically plausible but highly conservative samples. Statistical shape modeling generated broader morphometric variability while preserving plausibility and achieved the strongest downstream generalization at larger augmentation sizes. These findings show that downstream performance and geometric similarity alone are insufficient to characterize augmentation quality. The proposed framework provides an interpretable basis for selecting augmentation strategies according to application-specific objectives in small-sample biomedical imaging.