Low Field Pediatric Brain Magnetic Resonance Image Segmentation and Quality Assurance

Inhaltsverzeichnis

1. Frontmatter

   PDF-Version jetzt herunterladen

2. Task 1 - Automatic Ultra-Low Field MR Image Quality Assessment

   1. Frontmatter

   2. BRIQA: Balanced Reweighting in Image Quality Assessment of Pediatric Brain MRI

      • Open Access

      Alya Almsouti, Ainur Khamitova, Darya Taratynova, Mohammad Yaqub

      Abstract

      Assessing the severity of artifacts in pediatric brain Magnetic Resonance Imaging (MRI) is critical for diagnostic accuracy, especially in low-field systems where the signal-to-noise ratio is reduced. Manual quality assessment is time-consuming and subjective, motivating the need for robust automated solutions. In this work, we propose BRIQA (Balanced Reweighting in Image Quality Assessment), which addresses class imbalance in artifact severity levels. BRIQA uses gradient-based loss reweighting to dynamically adjust per-class contributions and employs a rotating batching scheme to ensure consistent exposure to underrepresented classes. Through experiments, no single architecture performs best across all artifact types, emphasizing the importance of architectural diversity. The rotating batching configuration improves performance across metrics by promoting balanced learning when combined with cross-entropy loss. BRIQA improves average macro F1 score from 0.659 to 0.706, with notable gains in Noise (0.430), Zipper (0.098), Positioning (0.097), Contrast (0.217), Motion (0.022), and Banding (0.012) artifact severity classification. The code is available at https://github.com/BioMedIA-MBZUAI/BRIQA.

      PDF-Version jetzt herunterladen

   3. Robust Multi-label Classification of MRI Artifacts in Low-Field Neonatal Brain Imaging via View-Conditional Dual-Task Learning

      • Open Access

      Cristian Lazo-Quispe, Roberto Espinoza-Chamorro

      Abstract

      Low-field MRI offers accessible neuroimaging in low-resource settings, but is often degraded by diverse artifacts that compromise diagnostic utility. In this work, we address Task 1 of the LISA Challenge 2025, which involves multi-label ordinal classification of seven artifact types in 3D neonatal brain MRIs acquired at 0.064T. Our pipeline employs a quality-aware 3D-to-2D projection strategy that automatically selects the optimal viewing plane based on voxel resolution, followed by view-conditional dual-task learning that jointly predicts artifact severity and brain bounding boxes. By combining brain-focused morphological preprocessing, MaxViT-based feature extraction with view embeddings, and multi-scale probability aggregation across slices, our approach achieves implicit spatial attention without explicit 3D modeling. To handle severe class imbalance and label ambiguity, we apply dynamic focal loss with class-specific weights, stratified patient-level cross-validation, and targeted data augmentation. By aggregating predictions across all valid 2D slices (80–120 per subject), we achieve a weighted F1 score of 0.771 on the test set. We analyze per-artifact performance and demonstrate that efficient 2D view-conditional modeling can match or exceed 3D approaches while maintaining computational efficiency suitable for clinical deployment in resource-limited settings.

      PDF-Version jetzt herunterladen

3. Task 2b - Automatic Basal Ganglia Segmentation from Ultra-Low Field MRI

   1. Frontmatter

   2. Towards Robust Basal Ganglia Segmentation in Ultra-Low-Field Pediatric MRI via an Optimized MS-TCNet

      • Open Access

      Yi Liu, Yueyue Zhu, Haotian Jiang, Xiaoyu Bai, Rongqing Cai, Geng Chen

      Abstract

      Magnetic Resonance Imaging (MRI) provides a non-invasive means to examine pediatric brain anatomy. However, in low-resource settings, ultra-low-field scanners are more widely used due to their affordability and portability, but they often generate images with a low signal-to-noise ratio and poor tissue contrast. This severely hampers the accurate delineation of critical subcortical structures such as the basal ganglia. In this work, we adapt and refine the multi-scale Transformer – CNN network (MS-TCNet) for bilateral basal ganglia segmentation in 0.064T pediatric MRI. Our optimized version, OMS-TCNet, integrates improved data augmentation and fine-tuned training configurations. Extensive experiments on the challenge dataset demonstrate that our method achieves robust and reliable segmentation performance under ultra-low-field imaging conditions. The code is publicly available at: https://github.com/Onion-Boy/OMS-TCNet.

      PDF-Version jetzt herunterladen

4. Tasks 2a and 2b - Automatic Hippocampal and Basal Ganglia Segmentation form Ultra-Low Field MRI

   1. Frontmatter

   2. Segmenting Brain Regions in Low Field Pediatric Brain MR Images Using (Symmetric) NnU-Net ResEnc

      • Open Access

      Jan Nikolas Morshuis, Matthias Hein, Christian F. Baumgartner

      Abstract

      The segmentation of low-field pediatric brain MR images is an important topic, as it can show the development of the pediatric brain. At the same time, the low cost and maintenance required by the low-field scanners make the technology more accessible in wider parts of the world compared to high-field MRI-scanners. The wider accessibility allows to understand the pediatric brain development in different regions of the world, thereby allowing to better understand the effects of, for example, nutrition for the brain development. Despite these advantages, automatically segmenting low-field MRI scans can be challenging: The internal brain structure can be hard to recognize in the low-field MRI images, the ground-truth segmentation can be unprecise due to registration errors between high-field and low-field MRI images, and the segmentation can potentially be different for the hippocampus region depending on whether it is the left or right hippocampus. In order to still be able to achieve the best possible segmentation scores, we try to increase the probability that our predicted segmentation is close to the ground-truth segmentation. To achieve this, we focus on extending the data-augmentation and make use of an ensemble of networks, of which we take the average as our final prediction. Doing so led to high-scores for the task 2 of the LISA challenge. The code is available at https://github.com/NikolasMorshuis/nnUNet-LISA-Challenge.git.

      PDF-Version jetzt herunterladen

   3. Segmenting Infant Brains Across Magnetic Fields: Domain Randomization and Annotation Curation in Ultra-low Field MRI

      • Open Access

      Vladyslav Zalevskyi, Dondu-Busra Bulut, Thomas Sanchez, Meritxell Bach Cuadra

      Abstract

      Early identification of neurodevelopmental disorders relies on accurate segmentation of brain structures in infancy, a task complicated by rapid brain growth, poor tissue contrast, and motion artifacts in pediatric MRI. These challenges are further exacerbated in ultra-low-field (ULF, 0.064 T) MRI, which, despite its lower image quality, offers an affordable, portable, and sedation-free alternative for use in low-resource settings. In this work, we propose a domain randomization (DR) framework to bridge the domain gap between high-field (HF) and ULF MRI in the context of the hippocampi and basal ganglia segmentation in the LISA challenge. We show that pre-training on whole-brain HF segmentations using DR significantly improves generalization to ULF data, and that careful curation of training labels, by removing misregistered HF-to-ULF annotations from training, further boosts performance. By fusing the predictions of several models through majority voting, we are able to achieve competitive performance. Our results demonstrate that combining robust augmentation with annotation quality control can enable accurate segmentation in ULF data. Our code is available at https://github.com/Medical-Image-Analysis-Laboratory/lisasegm

      PDF-Version jetzt herunterladen

   4. Enforcing Anatomical Symmetry with Euclidean Distance Transforms for Low-Field MRI Bilateral Structure Segmentation

      • Open Access

      Zdravko Marinov, Jens Kleesiek, Rainer Stiefelhagen

      Abstract

      Accurate segmentation of subcortical brain structures in MRI is essential for the study of neurodevelopment, particularly in pediatric populations. While low-field MRI scanners offer a cost-effective and safer alternative to high-field systems—especially eliminating the need for sedation in young children—they present challenges due to lower image resolution and signal-to-noise ratio. In this work, we propose a symmetry-aware post-processing strategy to improve the segmentation of bilateral structures in low-field MRI. We first train baseline U-Net models for the segmentation of eight anatomical structures, including hippocampi, in the LISA 2025 pediatric low-field MRI dataset. While these models achieve reasonable accuracy, we observe frequent violations of anatomical symmetry in their predictions. To address this, we introduce a novel correction step that explicitly enforces plausible anatomical symmetry by identifying discrepancies between hemispheres and applying deformation fields anchored by the dominant structure from each symmetric pair. This post-hoc alignment improves segmentation quality for all symmetric targets, particularly the hippocampi. Our approach highlights the importance of leveraging anatomical priors in low-resource imaging scenarios and paves the way for more reliable analyses in global health contexts(Code: https://github.com/Zrrr1997/LISA_2025_cvhci).

      PDF-Version jetzt herunterladen

   5. Coordinate Transformations Make Segmentation Models More Data-Efficient

      • Open Access

      Mahbod Issaiy

      Abstract

      Ultra-low-field (0.064T) magnetic resonance imaging (MRI) systems enable portable brain imaging but pose significant challenges for automated segmentation due to low signal-to-noise ratio and limited resolution. We present a coordinate transform-based deep learning approach for pediatric brain structure segmentation that analytically handles geometric variability through spherical and ellipsoidal coordinate mappings. Our method employs an ensemble of SwinUNETR models trained in Cartesian, spherical, and ellipsoidal spaces, combined with two novel loss functions: Projection Dice Loss for shape-aware supervision through 2D orthogonal projections, and Coordinate-Aware Soft Hausdorff Loss using coordinate-appropriate distance metrics. Evaluated on the LISA25 challenge dataset, our ensemble achieved competitive performance with Dice coefficients of 0.72±0.17 for hippocampus and 0.85±0.05 for basal ganglia segmentation. While coordinate transformations provide principled geometric handling, inverse transformation artifacts limited their individual effectiveness. The novel loss functions demonstrate clear benefits for medical image segmentation, advancing automated analysis capabilities for portable brain MRI systems in resource-constrained environments. Source code is available at https://github.com/mahbodez/LISA25-public.

      PDF-Version jetzt herunterladen

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

      • Open Access

      Kostiantyn Lavronenko, Rueveyda Yilmaz, Zhu Chen, Johannes Stegmaier, Volkmar Schulz

      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.

      PDF-Version jetzt herunterladen

   7. Automated Pediatric Brain Hippocampal and Basal Ganglia Segmentation in Ultra-low Field Magnetic Resonance Images

      • Open Access

      Toufiq Musah, Philip Nkwam, Ajay Sharma

      Abstract

      The automated segmentation of brain structures, such as the hippocampus and basal ganglia, from ultra-low field (0.064T) neonatal magnetic resonance imaging (MRI) is important for clinical diagnosis and neurodevelopmental research. Manual segmentation is time-consuming and subject to inter- and intra-observer variability. The unique challenges of ultra-low field MRI, including low signal to noise ratio and low spatial resolution make automated segmentation a difficult task. In this study, the challenge of automated segmentation of the basal ganglia and hippocampus in pediatric ultra-low field MRI is addressed. The approach builds on MedNeXt, a transformer-inspired, fully convolutional encoder-decoder architecture, trained with the nnU-Net pipeline. To address the class imbalance inherent in segmenting small structures, a combined Focal-Dice-CrossEntropy loss function was employed. The method was evaluated using the Dice Similarity Coefficient (DSC), 95th Percentile Hausdorff Distance (HD95), Average Symmetric Surface Distance (ASSD), and Relative Volume Error (RVE). The results show an average DSC of 0.69 ± 0.18 for hippocampal segmentation and an average DSC of 0.85 ± 0.06 for basal ganglia segmentation. The method demonstrated better performances in segmenting the basal ganglia in the ultra-low field images as compared to the hippocampus.

      PDF-Version jetzt herunterladen

5. Tasks 1, 2a and 2b Combined

   1. Frontmatter

   2. Application of Vision Transformers to Multi-task Learning in the LISA 2025 MRI Challenge

      • Open Access

      Tian Song, Dou Jiaqi

      Abstract

      Transformer-based architectures are increasingly used in medical image analysis to support diverse tasks under a unified framework. Our solution for the LISA 2025 challenge addresses both image quality classification (Task I) and semantic segmentation (Task II). For Task I, we introduced a slice-wise strategy based on a Vision Transformer (ViT) pre-trained on ImageNet. The model processes 3D MRI volumes decomposed into 2D slices, each carrying the original volume’s quality label. Predictions are combined by selecting the maximum value across slices for each quality category. The ViT encoder remained frozen throughout training, with updates limited to the classification layer. In Task II, a UNETR architecture was applied, incorporating encoder weights pre-trained on SAM-Med 3D. Training involved two stages: initial optimization of the decoder with a fixed encoder, followed by full model fine-tuning using Low-Rank Adaptation (LoRA). In the testing stage, Our approach achieved a weighted F1 score of 0.781 for quality assessment, and average Dice scores of 0.58 and 0.81 for hippocampal and basal ganglia segmentation, respectively. These outcomes highlight the flexibility and effectiveness of transformer-based models in multi-task medical image analysis. Our code for Task 1 has been made openly available at https://github.com/RimeT/lisa2025_task1_teamCGP.

      PDF-Version jetzt herunterladen

   3. Automatic Quality Assurance and Subcortical Brain Segmentation in Pediatric Ultra-Low-Field MRI: Exploring Ordinal Learning and Foundation Model Adaptation

      • Open Access

      Raquel González López, Maria Chiara Fiorentino, Gerard Martí-Juan, Oscar Camara, Miguel A. González Ballester

      Abstract

      Ultra-low-field (uLF) MRI systems offer portable and affordable neuroimaging solutions for pediatric patients and are valuable in resource-limited settings. However, such systems are susceptible to poor image quality, artifacts, and low contrast, making brain segmentation difficult. This study addresses two critical challenges in uLF MRI: automated quality assessment (QA) and anatomical structure segmentation. For QA, we propose a multi-label approach that incorporates the ordinal nature of artifact severity through an ordinal loss and models artifact co-occurrence patterns using Bayesian Networks. The approach is enhanced through aggressive synthetic data augmentation and ensemble learning, achieving a composite accuracy score of 0.84 across seven artifact categories. For segmentation, we benchmark a task-specific model (nnU-net) against a foundation model (SAM-Med3D) on the delineation of challenging subcortical structures. While nnU-Net, trained from scratch, achieved mean Dice score of 0.72 for hippocampi and 0.86 for basal ganglia, we demonstrate that lightweight fine-tuning of SAM-Med3D yields comparable results with a mean Dice score of 0.70 in hippocampi segmentation, despite domain shift. These results underscore the promise of foundation models for medical imaging in low-resource contexts, while highlighting the importance of domain adaptation. Overall, our pipeline represents a step forward in robust, automated QA and segmentation in uLF MRI for pediatric use. We release the code at https://github.com/reitxel/LISA2025TeamUPF.

      PDF-Version jetzt herunterladen

6. Backmatter

   PDF-Version jetzt herunterladen