Hardness-Aware Kernelized Contrastive Learning for Brain Age Prediction

Brain age estimation from structural magnetic resonance imaging is a valuable biomarker for characterizing deviations from normative aging, but robust prediction remains challenging under non-uniform age distributions and multi-site domain shifts. Existing contrastive regression methods typically assign pairwise weights according to chronological age differences, but they do not explicitly consider whether age-similar samples are properly aligned in the learned embedding space. This paper proposes a hardness-aware kernelized contrastive regression framework for brain age estimation from T1-weighted MRI. The proposed objective combines continuous label-aware kernel weighting with an adaptive hardness modulation term that emphasizes representation-label inconsistency, particularly age-similar pairs that remain distant in the embedding space. A 3D ResNet-18 encoder is trained with the proposed objective, and a ridge regression readout is used for age prediction from frozen embeddings. Experiments on the multi-site OpenBHB dataset and the ADNI dataset show consistent reductions in mean absolute error compared with MSE regression and representative contrastive regression baselines. The proposed method achieves MAEs of 3.84\((\pm)\)0.13 and 4.72\((\pm)\)0.17 on the OpenBHB internal and external splits, respectively, and 3.01\((\pm)\)0.06 on ADNI. Hyper-parameter analyses and representation visualizations suggest that hardness-aware weighting improves age-aligned embedding organization. These results indicate that modeling representation-label inconsistency is a useful strategy for continuous medical prediction, although further site-wise, bias-corrected, and clinically matched validation is required. The source code is available at https://anonymous.4open.science/r/hardness-aware-brain-age-A87B.