FDNet: A Lightweight Feature-Dominated U-Net for Hippocampus Segmentation in Brain MRI

Background/Objectives: Hippocampus segmentation in brain MRI is challenging because the structure is small, irregularly shaped, and poorly contrasted against surrounding tissue. Increasingly complex U-Net variants improve accuracy at computational cost. We investigated whether modifying the relative contribution of encoder features within skip connections can improve performance without additional attention or gating modules. Methods: We propose FDNet, a lightweight feature-dominated U-Net that concatenates multiple identical copies of encoder feature maps with the corresponding decoder features. Three variants (FDNetX2, FDNetX3, and FDNetX4) were evaluated on sagittal MRI slices from ADNI, the Medical Segmentation Decathlon, and IBSR using five-fold cross-validation, assessed with overlap metrics and per-slice false-positive/false-negative counts. To address potential subject-level leakage in the slice-level evaluation, a subject-grouped assessment with a fixed held-out test set was additionally conducted on ADNI, with re-implemented baselines and nnU-Net v2. Results: Under the slice-level protocol, FDNet achieved Dice coefficients of 0.9339 on ADNI, 0.9470 on Decathlon, and 0.9477 on IBSR, outperforming re-implemented U-Net, Attention U-Net, and SegNet baselines with substantially fewer parameters. Under the subject-grouped protocol, the best FDNet variant reached a held-out test Dice of 0.9050 versus 0.9019 for nnU-Net. All variants significantly outperformed the three baselines after Holm correction (adjusted p < 0.0001); their advantage over nnU-Net was not significant. Conclusions: Repeated encoder features within skip connections can provide competitive hippocampus segmentation with a relatively small parameter budget. Because all results come from curated research datasets without external or cross-dataset validation, they indicate competitive performance under the evaluated conditions rather than clinical readiness.