How reliable is deep learning segmentation of the infant brain? Variability and failure analysis of segmentation pipelines under a common protocol in 51 infants aged 1 to 9 months

Background Deep learning segmentation increasingly supplies the brain volumes used in infant neuroimaging, yet pipelines are usually ranked by a single mean Dice score, which hides variability and rare failures. Objective To quantify how much of the apparent difference between infant brain MRI segmentation pipelines reflects retraining variability, cross-validation partitioning, aggregation choice and rare failures, and to derive a reporting template for validating such tools. Materials and methods Four pipelines (nnU-Net, 3D TransUNet, Swin UNETR, 2.5D STSNet) and a cross-fitted consensus were trained and evaluated on the public Baby Open Brains dataset (71 T1 and T2-weighted MRI visits from 51 infants aged 1.0 to 9.4 months, 31 expert-corrected labels) with subject-disjoint five-fold cross-validation, subject-weighted summaries and Holm-corrected paired tests. Results Mean Dice was 0.8867 (95% confidence interval 0.8802–0.8932) for the consensus, 0.8805 for TransUNet, 0.8772 for Swin UNETR, 0.8722 for nnU-Net and 0.7868 for STSNet. Among the 3D pipelines the pipeline term explained 0.134% of Dice variation (𝑝 = 0.177), all paired effects were small (|𝑑| ≤ 0.22) and every mean-difference confidence interval included zero. TransUNet exceeded Swin UNETR in 46 of 51 infants; for nnU-Net versus Swin UNETR the mean (−0.0050) and median (+0.0072) disagreed in sign. Fold-level differences (−0.0490 to +0.0169) exceeded retraining standard deviations (0.0005–0.0016). Two scans drove the reversal and most failures; the consensus produced none. Conclusion Three competently configured 3D pipelines segmented the infant brain accurately with small mutual differences and no detected age dependence. Choosing among them mattered less than reporting variability and failures; a simple consensus removed all failures.