NEWS · RESEARCH · #357
Study finds sparse within-anatomy interpolation outperforms LightGBM cross-anatomy transfer for aortic FSI surrogates
Researchers evaluated a LightGBM geometry-only prior trained with leave-one-anatomy-out on three aortic geometries and zero-shot tested on a fourth, then applied sparse (5% anchor) field-completion; zero-shot transfer performed poorly, while within-anatomy interpolation methods (inverse-distance weighting and radial basis functions) achieved substantially higher R2 for OSI and stress metrics. Code, data and computation files are provided on GitHub and results are based on four de-identified human aortic models from the Vascular Model Repository (arXiv:2609.16322v1).
KEY POINTS
- Researchers evaluated a LightGBM geometry-only prior trained with leave-one-anatomy-out on three aortic geometries and zero-shot tested on a fourth, then applied sparse (5% anchor) field-completion; zero-shot transfer performed poorly, while within-anatomy interpolation methods (inverse-distance weighting and radial basis functions) achieved substantially higher R2 for OSI and stress metrics.
- Code, data and computation files are provided on GitHub and results are based on four de-identified human aortic models from the Vascular Model Repository (arXiv:2609.16322v1).
- This matters because it suggests that, for small cohorts, sparse within-anatomy labels and classic interpolation can outperform ML priors for filling biomechanics fields, tempering expectations for zero-shot cross-anatomy surrogate transfer.
WHY IT MATTERS
This matters because it suggests that, for small cohorts, sparse within-anatomy labels and classic interpolation can outperform ML priors for filling biomechanics fields, tempering expectations for zero-shot cross-anatomy surrogate transfer.