Yuyan Zhang
Anatomical priors are often assumed to improve medical image segmentation, but their role can be ambiguous when strong image evidence is available. We study this question in a controlled small-sample dual-modal brain MRI setting with 13-class coarse brain-region segmentation from T1 and FLAIR images. Using a quality-controlled 78-case subset and a subject-disjoint five-fold protocol, we compare compact U-Net variants, coordinate channels, modality dropout, fold-wise index-space atlas conditioning, and a full-to-missing modality consistency variant, FW-AtlasMC. Fold-wise probabilistic spatial priors are built only from training labels, and an atlas-only baseline quantifies spatial prior strength without model training. Under full T1+FLAIR input, learned methods are tightly clustered in Dice, and atlas-informed models do not significantly improve full-modality Dice over a 2.5D U-Net baseline. However, among learned image-conditioned models under missing-modality inference, FW-AtlasMC improves FLAIR-only Dice and reduces average performance drop compared with 2.5D U-Net, modality dropout, and FW-Atlas. FW-AtlasMC combines atlas-informed inputs with missing-modality training and full-to-missing consistency. Validation-only sensitivity analysis further indicates that the chosen consistency weight represents a reasonable trade-off within the tested range. These results suggest that anatomical priors are better understood as structural stabilizers under degraded modality information, rather than as universal full-modality accuracy boosters.