Localized Neighborhood Label Distribution Learning with Manifold-Regularization for Fetal Brain Age Estimation from MRI

Yiyuan Zhou, Ran Zhou, Wei Xia, Yu Guo, Zhongwei Huang, Haitao Gan

Abstract: Fetal brain age estimation is crucial for assessing neurodevelopment and detecting pathological changes of fetal. However, existing deep learning-based regression methods often neglect the label ambiguity introduced by developmental similarities between adjacent gestational weeks. To address this challenge, we first propose a localized neighborhood label distribution paradigm, which constrains the label distribution to the vicinity of the true gestational age, thus eliminating noise from distant labels. Second, we introduce a data-driven label distribution manifold regularization framework that hierarchically encodes both global and local label correlations. The global manifold enforces progressive dependencies across gestational weeks, suppressing the co-occurrence of non-adjacent gestational ages within the label distribution, while the local manifold optimizes stage-specific label associations based on MRI feature-driven clustering. Experimental results demonstrate that our framework outperforms existing methods in fetal brain age estimation, validating the efficacy of localized label distribution modeling and manifold regularization. This approach has the potential to assist clinicians in early detection and intervention for abnormal fetal brain development.

External IDs:dblp:conf/icic/ZhouZXGHG25