Abstract

Traditional evaluation metrics for medical image synthesis, such as Fréchet Inception Distance (FID), primarily assess global feature distributions but fail to guarantee the fine-grained anatomical and topological validity required for downstream medical imaging applications. While recent frameworks have introduced Wasserstein-based macroscopic shape comparisons to evaluate anatomical plausibility, they often overlook localized structural continuity. We introduce the Neuro-Fidelity Index (NFI), a multi-dimensional benchmarking framework for synthetic neuroimaging that integrates perceptual sharpness, microscopic topological continuity (Euler characteristics), and regional parcellation consistency into a unified metric using a weighted harmonic mean. The harmonic mean acts as an architectural bottleneck detector, ensuring that a catastrophic failure in any structural dimension is heavily penalized in the final score. We evaluate four representative generative architectures (VAE, HAGAN, LDM, and DDPM) on T1weighted brain MRIs from a cohort of 639 healthy adults. Through this benchmark, we identify a distinct “perceptual paradox” in latent-space modeling: while the LDM achieves high perceptual (0.887) and anatomical (0.867) similarity, it suffers from severe topological collapse (0.001) driven by geometric discontinuities introduced during latent decoding, resulting in an NFI of 0.002. In contrast, the pixel-space DDPM preserves structural integrity across all scales, achieving a balanced NFI of 0.833. These results demonstrate that the NFI serves as a rigorous, interpretable evaluation framework for medical image synthesis, revealing critical generative failures overlooked by conventional perceptual metrics.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/SASHIMI_028.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=l5qRzy408E

BibTex

@InProceedings{LeeWon_Beyond_MICCAISAT2026,
        author = { Lee, Won Hee AND Choi, Yongjin AND Kim, Yeonwoo AND Park, Hyeon Jun},
        title = { { Beyond Visual Realism: The Neuro-Fidelity Index for Validating Geometric Integrity in Synthetic Brain MRI } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17258},
        month = {pending},
        page = {pending}
}


back to top