[Paper] Deep infant brain segmentation from multi-contrast MRI

Source: arXiv - 2512.05114v1

Overview

The paper introduces BabySeg, a deep‑learning framework that can automatically segment infant and young‑child brain MRIs across a wide variety of scan protocols. By training the network on heavily randomized synthetic images, the authors achieve a single model that works reliably on different modalities, repeat scans, and even on image types it never saw during training—addressing a long‑standing bottleneck in pediatric neuro‑imaging.

Key Contributions

• Unified segmentation model that handles any combination of MRI contrasts (T1, T2, PD, etc.) without needing separate networks for each protocol.
• Domain‑randomization training pipeline that synthetically augments the data far beyond realistic bounds, making the model robust to scanner differences, motion artifacts, and missing modalities.
• Flexible feature‑pooling architecture that can ingest a variable number of input volumes and automatically learn how to fuse their information.
• State‑of‑the‑art accuracy on multiple public infant‑brain datasets, matching or surpassing specialized tools while running orders of magnitude faster.
• Open‑source implementation (including pretrained weights) that can be dropped into existing neuro‑imaging pipelines with minimal configuration.

Methodology

1. Data preparation – The authors collected a heterogeneous set of infant MRIs (ages 0–24 months) spanning several scanners, protocols, and quality levels.
2. Domain randomization – Instead of relying solely on real scans, they generated synthetic volumes by randomly varying contrast, intensity scaling, noise levels, bias fields, and even adding unrealistic artifacts. This forces the network to learn features that are invariant to such shifts.
3. Multi‑input encoder – Each input scan passes through a shallow encoder; the resulting feature maps are concatenated and processed by a shared decoder. Because the concatenation operates on a list, the network can accept 1, 2, … N scans at inference time.
4. Training objective – A standard Dice loss is combined with a boundary‑aware term to sharpen cortical surface predictions.
5. Implementation details – The backbone is a 3‑D U‑Net variant with group normalization, trained on 8‑GPU clusters for ~48 h. The final model size is ~120 MB, enabling deployment on workstation‑class CPUs/GPUs.

Results & Findings

• Accuracy: BabySeg consistently outperformed or matched dedicated tools such as iBEAT, MANTiS, and infant‑FreeSurfer across all age brackets.
• Robustness: When one modality was deliberately omitted (e.g., only T1 available for a model trained with T1+T2), performance degraded only marginally, confirming the flexible pooling design.
• Speed: Inference takes < 2 seconds per volume on a modern GPU, compared to several minutes for traditional atlas‑based pipelines.

Practical Implications

• Clinical workflow integration – Radiology departments can run BabySeg on routine neonatal scans without worrying about missing contrasts, enabling near‑real‑time brain volumetrics for early diagnosis of developmental disorders.
• Research scalability – Large longitudinal studies (e.g., tracking brain growth from birth to 2 years) can process thousands of scans in a fraction of the time previously required, freeing up compute resources for downstream analyses.
• Cross‑site studies – Because the model tolerates scanner‑specific quirks and motion artifacts, multi‑center collaborations no longer need extensive harmonization pipelines.
• Tooling ecosystem – The open‑source code can be wrapped into popular neuro‑imaging frameworks (Nipype, BIDS‑Apps), allowing developers to plug BabySeg into existing pipelines with a single command.

Limitations & Future Work

• Training data bias – Although domain randomization mitigates many shifts, the underlying real‑world data still skew toward high‑resource hospitals; performance on ultra‑low‑field or highly noisy bedside scanners remains untested.
• Age‑specific anatomy – The model treats age as an implicit factor; explicit age conditioning could further improve segmentation of rapidly changing structures (e.g., myelination fronts).
• Extension to pathology – Current evaluation focuses on typical development; future work should assess robustness on infants with brain injuries, hydrocephalus, or congenital malformations.
• Explainability – Providing uncertainty maps or attention visualizations would help clinicians trust automated segmentations in borderline cases.

Bottom line: BabySeg demonstrates that a single, cleverly trained deep network can replace a zoo of specialized infant‑brain segmentation tools, offering developers and clinicians a fast, robust, and easy‑to‑integrate solution for pediatric neuro‑imaging.

Authors

• Malte Hoffmann
• Lilla Zöllei
• Adrian V. Dalca

Paper Information

• arXiv ID: 2512.05114v1
• Categories: cs.LG, cs.CV, eess.IV
• Published: December 4, 2025
• PDF: Download PDF