Atlas-Augmented Semantic Segmentation for Robust Ultra-Low-Field Pediatric Brain Imaging

Kostiantyn Lavronenko, Rueveyda Yilmaz, Zhu Chen, Johannes Stegmaier, Volkmar SchulzLecture Notes in Computer Science 2026 · DOI · OpenAlex

Abstract

Low-field MRI offers a portable, cost-effective alternative to conventional high-field scanners but suffers from reduced signal-to-noise ratio and spatial inhomogeneity, which compromise the accuracy and consistency of automated brain structure segmentation. In this work, we introduce atlas-augmented deep learning models that integrate probabilistic anatomical priors to enhance the delineation of pediatric hippocampus and basal ganglia in ultra-low-field MRI (0.064 T). We evaluate seven pipelines on the LISA 2025 dataset (79 T2-weighted scans): baseline VNet, nnU-Net, and MedSAM2 variants (2D and 3D decoders), as well as atlas-augmented VNet, atlas-augmented nnU-Net, and atlas-augmented MedSAM2-3D. For VNet and MedSAM2-3D, probabilistic maps from the Pauli and Harvard-Oxford atlases are encoded and fused with intermediate feature maps, while nnU-Net ingests priors as additional input channels. Baseline nnU-Net attains mean DSCs of 0.71 for hippocampus and 0.86 for basal ganglia; atlas augmentation yields modest hippocampal gains (HD95 $\downarrow$ 0.05, ASSD $\downarrow$ 0.06) and more pronounced improvements in basal ganglia segmentation, reflecting richer prior information for larger structures. VNet and MedSAM2 variants exhibit limited hippocampal benefit, highlighting the strength of nnU-Net’s adaptive framework. Our findings establish atlas-augmented nnU-Net as a new benchmark for robust segmentation in resource-constrained, low-field imaging environments.

Introduction

Ultra-low-field (uLF) MRI scanners (0.064 T) represent a significant advancement in global health, offering portable, cost-effective, and safe imaging for pediatric patients. However, these scanners produce images with low signal-to-noise ratios (SNR) and significant spatial inhomogeneity, making automated segmentation of brain structures exceptionally challenging. Pediatric brains further complicate this due to lower contrast between gray and white matter compared to adults. This research addresses these challenges by augmenting deep learning models with anatomical "atlases"—probabilistic anatomical priors that serve as a stabilizer during the segmentation process in noisy environments.

Methods

The study evaluates seven pipelines based on three core architectures: VNet, nnU-Net, and MedSAM2. The primary innovation is Atlas Augmentation, which integrates probabilistic maps from the Pauli and Harvard-Oxford atlases to provide the model with anatomical context.

Overview of the proposed atlas-augmented architectures for nnU-Net, MedSAM2, and VNet.

Two integration strategies were tested:

  1. Feature Fusion (VNet and MedSAM2): A dedicated "Atlas Encoder" processes the priors, which are then fused with image features deep within the network.
  2. Input Concatenation (nnU-Net): Atlas maps are appended as additional input channels, allowing nnU-Net’s adaptive preprocessing to handle the integration.

Dataset and Benchmark

The models were trained and validated using the LISA 2025 dataset, comprising 79 volumetric T2-weighted pediatric brain scans acquired on a Hyperfine 0.064T system. Ground truth labels were generated by expert neuroimagers on high-field scans and coregistered to the uLF images to ensure high-quality training data.

Results and Insights

The atlas-augmented nnU-Net [nnU-Net(A)] emerged as the top-performing model, establishing a new benchmark for uLF segmentation. It achieved a mean Dice Similarity Coefficient (DSC) of 0.71 for the hippocampus and 0.86 for the basal ganglia.

Qualitative comparison of ground truth (GT) and model predictions for the hippocampus and basal ganglia.

Key findings from the evaluation include:

  • The Power of Priors: Anatomical atlases act as a stabilizer. When the image is too noisy to clearly define a boundary, the atlas provides a "best guess" based on human anatomy, preventing significant errors.
  • Architecture Robustness: Despite the rise of foundation models like MedSAM2, nnU-Net’s self-configuring framework proved most effective. Its ability to automatically adapt preprocessing to the specific dataset was more valuable for uLF MRI than broad pretraining.
  • Structure Size Sensitivity: Larger, more stable structures like the basal ganglia benefited more from atlas augmentation than smaller, more variable structures like the hippocampus, as they align more consistently with global anatomical priors.

Discussion and Future Work

While atlas augmentation significantly improves robustness, the approach has limitations. The study utilized atlases derived from older cohorts; using age-matched pediatric templates could further improve accuracy for infants. Additionally, the computational cost of registering a patient's scan to an atlas remains a hurdle for real-time clinical use. Future work will explore parameter-efficient fine-tuning (e.g., LoRA) to better adapt large foundation models to the unique noise profiles of ultra-low-field hardware.

Summary updated

Abstract

Low-field MRI offers a portable, cost-effective alternative to conventional high-field scanners but suffers from reduced signal-to-noise ratio and spatial inhomogeneity, which compromise the accuracy and consistency of automated brain structure segmentation. In this work, we introduce atlas-augmented deep learning models that integrate probabilistic anatomical priors to enhance the delineation of pediatric hippocampus and basal ganglia in ultra-low-field MRI (0.064 T). We evaluate seven pipelines on the LISA 2025 dataset (79 T2-weighted scans): baseline VNet, nnU-Net, and MedSAM2 variants (2D and 3D decoders), as well as atlas-augmented VNet, atlas-augmented nnU-Net, and atlas-augmented MedSAM2-3D. For VNet and MedSAM2-3D, probabilistic maps from the Pauli and Harvard-Oxford atlases are encoded and fused with intermediate feature maps, while nnU-Net ingests priors as additional input channels. Baseline nnU-Net attains mean DSCs of 0.71 for hippocampus and 0.86 for basal ganglia; atlas augmentation yields modest hippocampal gains (HD95 $$\downarrow $$ 0.05, ASSD $$\downarrow $$ 0.06) and more pronounced improvements in basal ganglia segmentation, reflecting richer prior information for larger structures. VNet and MedSAM2 variants exhibit limited hippocampal benefit, highlighting the strength of nnU-Net’s adaptive framework. Our findings establish atlas-augmented nnU-Net as a new benchmark for robust segmentation in resource-constrained, low-field imaging environments. The code for our methods will be publicly accessible after the successful publication of the paper here: https://github.com/mackostya/deepatlas-ulf-seg .
BibTeX
@inproceedings{lavronenko2026atlas,
  author = {Kostiantyn Lavronenko and Rueveyda Yilmaz and Zhu Chen and Johannes Stegmaier and Volkmar Schulz},
  title = {{Atlas-Augmented Semantic Segmentation for Robust Ultra-Low-Field Pediatric Brain Imaging}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2026},
  doi = {10.1007/978-3-032-14417-1\_8},
}
Figure