Medical Image Analysis

Accurate measurement of intra-cardiac blood oxygen (O2) saturation is essential for cardiovascular assessment, yet current methods require invasive catheterization. T2-based cardiac magnetic resonance imaging (CMRI) enables non-invasive O2 quantification, but deep learning automation is constrained by scarce annotated data. We propose a unified self-supervised learning (SSL) framework integrating cine CMRI and T2 oximetry CMRI to learn generalizable representations without labels. Our approach pre-trains ResNet and vision transformer encoders using contrastive learning and masked image modeling on over 48,000 cardiac images. Pre-trained encoders are fine-tuned for O2 saturation regression with uncertainty quantification to enhance clinical trustworthiness. Our SSL framework significantly outperforms traditional radiomics and supervised baselines, with SimCLR pre-trained ResNet achieving a mean absolute error of 3.70, representing over 15\% improvement. These findings demonstrate SSL's potential to address annotation bottlenecks in non-invasive cardiac diagnostics.

Background: Segmentation of the left ventricular myocardium, left ventricular cavity, and right ventricular cavity on short-axis cine cardiac magnetic resonance (CMR) images is essential for quantifying cardiac structure and function. However, existing automated segmentation tools are limited by small training datasets, narrow disease coverage, restrictive input format requirements, and the absence of anatomical plausibility constraints, hindering their clinical adoption. Methods: We constructed the largest annotated CMR short-axis segmentation dataset to date, comprising 1,555 subjects from 12 centers with five cardiac disease types and full cardiac cycle annotations totaling 319,175 labeled images. A MedNeXt-L model was trained using a 2D slice-by-slice strategy with full field-of-view input, eliminating dependencies on 3D volumes, temporal sequences, or region-of-interest(ROI) localization. A deterministic three-step post-processing pipeline was designed to enforce anatomical priors: connected component constraint, containment relationship constraint, and gap-filling constraint. The model was validated on an internal test set (310 subjects) and three independent public external datasets (ACDC, M&Ms1, M and Ms2; 855 subjects from 6 additional centers across 3 countries), spanning 15 cardiac disease categories-10 of which were never encountered during training. Results: The model achieved mean Dice similarity coefficients (DSC) of 0.913 {+/-} 0.037 and 0.911 {+/-} 0.040 on internal and external test sets, respectively, with a cross-domain performance gap of only 0.002. Post-processing eliminated all containment violations (7.5% [->] 0%) and gap errors (1.8% [->] 0%) while reducing fragment rates by 85.5% (9.0% [->] 1.3%). Zero-shot generalization to 10 unseen disease categories yielded DSC values ranging from 0.899 to 0.921. Automated clinical functional parameters demonstrated excellent agreement with manual measurements for left ventricular indices and right ventricular volumes (intraclass correlation coefficients [≥] 0.977). Conclusions: CorSeg-CineSAX provides a robust, open-source framework for fully automatic CMR short-axis segmentation across diverse clinical scenarios. All source code and pre-trained weights are publicly available at https://github.com/RunhaoXu2003/CorSeg.

Deep segmentation networks can degrade sharply when an expected MRI sequence is unavailable at inference. We present NeuroMesh, a bottleneck controller that combines a gated recurrent unit (GRU) with a graphconvolutional edge-activation mask, designed to adapt a U-Net-style segmentation backbone to missing input. We evaluate NeuroMesh in a pilot study using a 30-patient subset of the BraTS 2020 benchmark (22 training, 4 validation, and 4 held-out test patients) under a prespecified frozentest protocol. On the frozen test set, NeuroMesh has higher tumor-core and enhancing-tumor Dice than a plain U-Net in most evaluated missing-modality conditions, but wholetumor Dice falls from 0.596 to 0.108 when FLAIR is missing, compared with 0.604 to 0.545 for the plain U-Net. Direct analysis of the predicted edge-activation mask shows negligible change across modality-availability conditions. A parameter-light static-gating control reproduces the FLAIR failure mode without recurrence, a failure-signal input, or graph-structured machinery. These results do not support the intended interpretation that the trained controller performs input-conditional topology rewiring at the scale of this pilot. Instead, they expose a discrepancy between architectural intent and realized behavior and identify a specific missing-modality failure mode that warrants further investigation. Given the small validation and test sets, the findings are descriptive and do not establish clinical or population-level generalization.

Accurate rectal tumor segmentation using magnetic resonance imaging (MRI) is paramount for effective treatment planning. It allows for volumetric and other quantitative tumor assessments, potentially aiding in prognostication and treatment response evaluation. Manual delineation of rectal tumors and surrounding structures is time-consuming and typically. Over the past few years, deep learning has shown strong results in automated tumor segmentation in MRI. Current studies on automated rectal tumor segmentation, however, focus solely on tumoral regions without considering the rectal anatomical entities and often lack a solid multicenter external validation. In this study, we improved rectal tumor segmentation by incorporating anomaly maps derived from anatomical inpainting. This inpainting was implemented using a U-Net-based model trained to reconstruct a healthy rectum and mesorectum from prostate T2-weighted images (T2WI). The rectal anomaly maps were generated from the difference between the original rectal and reconstructed pseudo-healthy slices during inference. The derived anomaly maps were used in the downstream tumor segmentation tasks by fusing them as an additional input channel (AAnnUNet). Alternative methods for integrating rectal anatomical knowledge were evaluated as baselines, including Multi-Target nnUNet (MTnnUNet), which added rectum and mesorectum segmentation as auxiliary tasks, and Multi-Channel nnUNet (MCnnUNet), which utilized rectum and mesorectum masks as an additional input channel. As part of this study, we benchmarked nine models for rectal tumor segmentation on a large multicenter dataset of preoperative T2WI as the baseline and nnUNet outperformed the other eight models on the external dataset. The MTnnUNet demonstrated improvements in both supervised and semi-supervised settings (AI-generated rectum and mesoretum were used) compared to nnUNet, while the MCnnUNet showed benefits only in the semi-supervised setting. Importantly, anomaly maps were strongly associated with tumoral regions, and their integration within AAnnUNet led to the best tumor segmentation results across both settings. The effectiveness of AAnnUNet demonstrated the value of the anomaly maps, indicating a promising direction for improving rectal tumor segmentation and model robustness for multicenter data.

With the increasing ability to record neuroimaging with higher spatial and temporal resolution, there is a growing need for methods that reduce these high-dimensional representations into latent low-dimensional structures that are discriminative and/or predictive of behavior, disease, or in general experimental context. We propose a metric learning framework to extract meaningful latent structures from high-dimensional fMRI data. This method learns the latent embeddings that reduce the intra-group variability while maximizing the inter-group variability. In addition, our method leverages advances in few-shot learning approaches to adapt to small sample-size fMRI datasets, allowing one to learn the latent structure from just a few samples per context. We evaluate our work on two publicly available fMRI datasets and report superior results compared to popular alternative approaches such as Principal Component Analysis (84.7% vs. 60%; 21.8% vs. 8.3%). We provide the Python code as open-source at Github.

Congenital heart disease (CHD) is the most commonly diagnosed birth defect. T2w black blood MRI provides optimal vessel visualisation, aiding prenatal CHD diagnosis. Common clinical practice involves manual segmentation of fetal heart and vessels for visualisation and reporting purposes. We propose an automated multi-label fetal cardiac vessels deep learning segmentation approach for T2w black blood MRI. Our network is trained using single-label manual segmentations obtained through current clinical practice, combined with a multi-label anatomical atlas with desired multi-label segmentation protocol. Our framework combines deep learning label propagation with 3D residual U-Net segmentation to produce high-quality multi-label output well adapted to the individual subject anatomy. We train and evaluate the network using forty fetal subjects with suspected coarctation of the aorta, achieving a dice score of 0.79 {+/-} 0.02 for the fetal cardiac vessels region. The proposed network outperforms the label propagation and achieves a statistically equivalent performance to a 3D residual U-Net trained exclusively on manual single-label data (p-value>0.05). This multi-label framework therefore represents an advancement over the single-label approach, providing label-specific anatomical information, particularly useful for assessing specific anomaly areas in CHD.

Computational pathology is revolutionizing the field of pathology by integrating advanced computer vision and machine learning technologies into diagnostic workflows. It offers unprecedented opportunities for improved efficiency in treatment decisions by allowing pathologists to achieve higher precision and objectivity in disease classification, tumor microenvironment description and identification of new biomarkers. However, the potential of computational pathology in personalized medicine comes with significant challenges, particularly in annotating whole slide images (WSI), which is time-consuming, costly and subject to inter-observer variability. To address these challenges, Self-Supervised Learning (SSL) has emerged as a promising solution to learn representations from histology patches and leverage large volumes of unlabelled WSI. Recently, Masked Image Modeling (MIM) as a SSL framework has emerged and is now considered to outperform purely contrastive learning paradigms. In this work, we therefore explore the application of MIM to histology using iBOT, a self-supervised transformer-based framework. Through a wide range of 17 downstream tasks over seven cancer indications, both at the slide and patch levels, we provide recommendations on the pre-training of large models for histology data using MIM. First, we demonstrate that in-domain pre-training with iBOT outperforms both ImageNet pre-training and a model pre-trained with a purely contrastive learning objective, MoCo v2. Second, we show that Vision Transformers (ViT) models, when scaled appropriately, have the capability to learn pan-cancer representations that benefit a large variety of downstream tasks. Finally, our iBOT ViT-Base model (80 million parameters), pre-trained on more than 40 million histology images from 16 different cancer types, achieves state-of-the-art performance in most weakly-supervised WSI classification tasks compared to other SSL frameworks available in the literature. This paves the way for the development of a foundation model for histopathology. Our code, models and features are publicly available at https://github.com/owkin/HistoSSLscaling.

Mapping hierarchical brain networks within traditional Euclidean space causes significant structural distortion, undermining neuroimaging diagnostic frameworks. While hyperbolic models like the Poincare ball preserve these nested topologies, they demand heavy computational overhead due to intricate Mobius operations and curved geodesics. This paper introduces a highly efficient non-Euclidean framework for analyzing neurocognitive decline utilizing the Beltrami-Klein ball model. By projecting hyperbolic geodesics as Euclidean straight lines, this approach converts complex distance calculations into simple dot products, radically reducing processing demands. We validated our methodology against state-of-the-art Poincare and Lorentz baselines using datasets for Schizophrenia, Parkinsons Disease, and Alzheimers Disease. The Klein-based framework demonstrates superior performance, delivering both higher diagnostic precision and accelerated processing velocities across all three neurocognitive disorders.

Deep learning (DL)-based image synthesis has recently gained enormous interest in medical imaging, allowing for generating multi-contrast data and therefore, the recovery of missing samples from interrupted or artefact-distorted acquisitions. However, the accuracy of DL models heavily relies on the representativeness of the training datasets naturally characterized by their distributions, experimental setups or preprocessing schemes. These complicate generalizing DL models across multi-site heterogeneous data sets while maintaining the confidentiality of the data. One of the possible solutions is to employ federated learning (FL), which enables the collaborative training of a DL model in a decentralized manner, demanding the involved sites to share only the characteristics of the models without transferring their sensitive medical data. The paper presents a DL-based magnetic resonance (MR) data translation in a FL way. We introduce a new aggregation strategy called FedBAdam that couples two state-of-the-art methods with complementary strengths by incorporating momentum in the aggregation scheme and skipping the batch normalization layers. The work comprehensively validates 10 FL-based strategies for an image-to-image multi-contrast MR translation, considering healthy and tumorous brain scans from five different institutions. Our study has revealed that the FedBAdam shows superior results in terms of mean squared error and structural similarity index over personalized methods, like the FedMRI, and standard FL-based aggregation techniques, such as the FedAvg or FedProx, considering multi-site multi-vendor heterogeneous environment. The FedBAdam has prevented the overfitting of the model and gradually reached the optimal model parameters, exhibiting no oscillations.

IntroductionCerebral amyloid angiopathy (CAA) is characterized by amyloid-beta deposition in cortical and leptomeningeal vessels and associated with cognitive impairment and hemorrhage. Current neuropathological assessments rely on semiquantitative grading and lack vessel-level resolution and scalability. Existing computational pathology approaches also fail to capture individual vessel morphology and spatial amyloid distribution across whole-slide images (WSIs). To address this gap, we developed a deep learning framework for reproducible, quantitative analysis of CAA in WSIs. MethodsWe analyzed 20 postmortem brain tissue sections from the frontal (n = 10) and occipital cortices (n = 10) of 10 individuals with Alzheimers disease pathology obtained from the University of Pittsburgh Alzheimers Disease Research Center, which served as the internal development cohort. An independent external cohort consisted of 10 sections (5 frontal and 5 occipital samples) from 5 individuals obtained from the University of Kentucky Alzheimers Disease Research Center. We trained and compared three semantic segmentation architectures, a standard U-Net, a dual-attention residual U-Net (DA-ResUNet), and a Swin Transformer-based U-Net (Swin-UNet), using the internal development cohort with slide-level five-fold cross-validation. All models were evaluated on the independent external cohort to assess generalization under domain shift. Based on segmentation performance and computational efficiency, we selected one architecture to generate whole-slide composite segmentation masks for vessel walls, amyloid deposits, and tissue compartments. These masks were subsequently used for deterministic vessel detection, morphometric measurements, and quantification of vascular and perivascular amyloid features through post-processing analysis. ResultsAll three architectures achieved high segmentation accuracy on the internal cohort, with Dice scores above 90% across vessel walls, amyloid deposits, gray matter, and leptomeninges. The Swin-UNet showed marginally higher performance for vessel segmentation, whereas the DA-ResUNet provided more balanced accuracy and computational efficiency and was selected for downstream analysis. External cohort evaluation demonstrated robust generalization, with attention-enhanced models outperforming the standard U-Net under domain shift. Using the selected model, the pipeline reliably detected valid vessels, excluded non-vascular artifacts, and enabled deterministic extraction of vessel morphometry, vascular and perivascular amyloid burden, and identification of circumferential CAA involvement at the vessel level. DiscussionThis framework provides a scalable, interpretable solution for vessel-level CAA analysis, supporting robust geometric and spatial characterization of cerebrovascular pathology and enabling future integration with clinical and genetic studies. Beyond CAA, the modular design allows extension to other vascular pathologies, including arteriolosclerosis, in WSIs, facilitating broader investigation of cerebrovascular disease mechanisms.

Large data initiatives and high-powered brain imaging analyses require the pooling of MR images acquired across multiple scanners, often using different protocols. Prospective cross-site harmonization often involves the use of a phantom or traveling subjects. However, as more datasets are becoming publicly available, there is a growing need for retrospective harmonization, pooling data from sites not originally coordinated together. Several retrospective harmonization techniques have shown promise in removing cross-site image variation. However, most unsupervised methods cannot distinguish between image-acquisition based variability and cross-site population variability, so they require that datasets contain subjects or patient groups with similar clinical or demographic information. To overcome this limitation, we consider cross-site MRI image harmonization as a style transfer problem rather than a domain transfer problem. Using a fully unsupervised deep-learning framework based on a generative adversarial network (GAN), we show that MR images can be harmonized by inserting the style information encoded from a reference image directly, without knowing their site/scanner labels a priori. We trained our model using data from five large-scale multi-site datasets with varied demographics. Results demonstrated that our styleencoding model can harmonize MR images, and match intensity profiles, successfully, without relying on traveling subjects. This model also avoids the need to control for clinical, diagnostic, or demographic information. Moreover, we further demonstrated that if we included diverse enough images into the training set, our method successfully harmonized MR images collected from unseen scanners and protocols, suggesting a promising novel tool for ongoing collaborative studies.

Histological staining is vital in clinical pathology for visualizing tissue structures. However, these techniques are laborious and time-consuming. Digital virtual staining offers a promising solution, but existing methods typically rely on Generative Adversarial Networks (GANs), which may suffer from artifacts and mode collapse. Motivated by the success of diffusion models, we present DUST, a novel Diffusion-based Unified framework for versatile Stain Transfer in histopathology. To enhance domain awareness and task-specific performance, we propose a dual encoding strategy that integrates the stain types of both the source and target domains. Additionally, we introduce a dynamic dual-output head to address the unstable intensity issue encountered with conventional DDPM implementations. Validated on a curated fourstain kidney histopathological dataset (H&E, MT, PAS, and PASM), DUST demonstrates superior versatile stain transfer capabilities. Our research highlights the potential of diffusion models to advance virtual staining, paving the way for more efficient digital pathology analyses.

In medical image classification tasks like the detection of diabetic retinopathy from retinal fundus images, it is highly desirable to get visual explanations for the decisions of black-box deep neural networks (DNNs). However, gradient-based saliency methods often fail to highlight the diseased image regions reliably. On the other hand, adversarially robust models have more interpretable gradients than plain models but suffer typically from a significant drop in accuracy, which is unacceptable for clinical practice. Here, we show that one can get the best of both worlds by ensembling a plain and an adversarially robust model: maintaining high accuracy but having improved visual explanations. Also, our ensemble produces meaningful visual counterfactuals which are complementary to existing saliency-based techniques. Code is available under https://github.com/valentyn1boreiko/Fundus_VCEs.

In studying the brains functional connectivity and its associations with clinically observed assessments, novel learning frameworks modeling its network properties in conjunction with assessment variables are crucial to uncover variable-specific patterns via meaningful encoding and reconstruction. We present a generative framework for modeling human brain functional connectivity features while retaining key network metrics and differences associated with demographic and cognitive variables. A conditional graph variational autoencoder is employed to encode static functional network connectivity (sFNC) features into a latent representation, which is then utilized for the dual purpose of reconstructing sFNC data conditioned on variables such as biological sex or fluid intelligence, and identifying discriminative connectivity features associated with the conditioning variables in the latent space. Using over 20,000 subjects from the UK Biobank, our model demonstrates high-fidelity reconstructions that preserve condition-specific network patterns, while the latent space captures interpretable patterns associated with these variables. The group differences in latent space are highlighted by one-hot probing of the latent dimensions and forward mapping to connectivity patterns. This approach provides a scalable, network-informed framework for studying brain functional connectivity and its associations with individual differences, offering potential applications in characterizing functional signatures for mental health conditions via clinically observed assessment variables. AUTHOR SUMMARYTo enable the modeling of the brain functional connectivity network for encoding and reconstructing assessment-specific differences, we propose a conditional graph-based generative framework for modeling human brain functional connectivity while accounting for demographic and cognitive differences. Using a conditional graph variational autoencoder, our approach learns interpretable latent representations of functional connectivity networks derived from fMRI data. Evaluated on over 20,000 UK Biobank subjects, the model accurately reconstructs connectivity patterns outperforming baseline architectures and preserves differences associated with biological sex and fluid intelligence. By probing the latent space and mapping latent dimensions back to brain networks, we identify condition-specific connectivity features in an interpretable manner. This work provides a scalable, network-informed approach for studying individual differences in functional brain organization.

The function of the human heart is characterised by complex patterns of motion that change throughout our lifespan due to accumulated damage across biological scales. Understanding the drivers of cardiac ageing is key to developing strategies for attenuating age-related processes. The motion of the surface of the heart can be conceived as a graph of connected points in space moving through time. Here we develop a generalisable framework for modelling three-dimensional motion as a graph and apply it to a task of predicting biological age. Using sequences of segmented cardiac imaging from 5064 participants in UK Biobank we train a graph neural network (GNN) to learn motion traits that predict healthy ageing. The GNN outperformed (mean absolute error, MAE = 4.74 years) a comparator dense neural network and boosting methods (MAE = 4.90 years and 5.08 years, respectively). We produce human-intelligible explanations of the predictions and using the trained model we also assess the effect of hypertension on biological age. This work shows how graph representations of complex motion can efficiently predict biologically meaningful outcomes.

BackgroundSuper-resolution reconstruction (SRR) of fetal brain magnetic resonance imaging has the potential to enable the development of new imaging biomarkers to better study in utero neurodevelopment. However, potential biases in 2D biometric and 3D volumetric measurements due to different SRR techniques remain understudied. PurposeTo assess the consistency of biometric and volumetric measurements across three hospitals using three widely used SRR pipelines. Materials and MethodsThis retrospective study used T2-weighted (T2w) fetal brain MRI scans acquired in routine clinical practice at three hospitals. MRIs from each subject were reconstructed with each of the 3 SRR methods. Four experts did biometric measurements on each SRR volume blinded to the method used. Automated 3D volumetry was performed using a state-of-the-art segmentation method. A univariate analysis was first carried out with Friedman tests with post-hoc Wilcoxon rank-sum tests, and results were confirmed in a multivariate analysis accounting for the effect of gestational age and different raters, using a t-distributed generalized additive model. An additional qualitative evaluation was performed to assess how likely clinicians would be to use the current SRR volumes in their practice, and whether they would prefer it to low-resolution T2w acquisitions. Differences were assessed with Friedman tests and post-hoc Wilcoxon rank-sum tests. Results84 healthy subjects were included in three gestational age groups ([21-28): 25.4{+/-}1.9, [28-32): 29.3{+/-}1.3, [32-36): 33.5{+/-}1.2). Statistically significant differences in biometric measurements were found, but consistently remained below voxel width (0.8 mm). Automated 3D volumetry revealed systematic but very small effects (<2.8%). The qualitative evaluation showed systematic differences between SRR methods for the perception of white matter intensity (p=0.02) and sharpness of the image (p=0.01). ConclusionVariations in 2D and 3D quantitative measurements did not show any large systematic bias when using different SRR methods for radiological assessment in clinical routine across multiple centers, scanners, and raters. SummaryDifferent super-resolution reconstruction methods for fetal brain MRI volumes lead to negligible variations in 2D or 3D quantitative measurements; this may help achieve larger sample sizes in prenatal development studies. Key Results- In this multi-centric retrospective study, 252 super-resolution reconstructions (SRR) scans from 84 healthy subjects showed negligible variations in 2D in biometric measures (below the voxel with of 0.8 mm; p<0.001). - 3D measurements revealed small variations ranging from 0.8 % in supratentorial tissues (p<0.001) to 2.8% in the extra-cerebral cerebrospinal fluid (p<0.001). - Clinicians favored having both low resolution and SRR volumes available.

Whole slide image (WSI) classification is crucial in computational pathology, yet the gigapixel scale of WSIs makes it challenging to extract discriminative and compact WSI-level features for disease diagnosis. In this paper, we propose MambaWSI, a novel method that leverages the state space model (SSM) for WSI classification by exploring multi-scale local and global features. Unlike existing approaches that sequentially traverse WSI tiles and rely on vanilla SSMs for long-range dependency modeling, we exploit a traversal strategy in a higher-dimensional discrete space that preserves spatial proximity, enabling a first-local-then-global feature extraction process. Furthermore, to align with the clinical workflow of pathologists when examining WSIs at multiple scales, we propose a two-stage hierarchical fusion strategy: inter-scale feature alignment and aggregation, followed by attention-based fusion across magnifications, integrating complementary information from multiple magnifications. Experiments on two datasets demonstrate that MambaWSI outperforms state-of-the-art methods in classification performance.1

Accurate brain tumor segmentation in magnetic resonance imaging (MRI) remains a critical challenge due to complex tumor heterogeneity, fuzzy boundaries, and significant inter-patient variability. In this study, we propose Diet-Seg (Difficulty-Informed Edge-enhanced Tiny Segmentation), a novel segmentation framework that integrates entropy-based pixel-wise hardness estimation into the training process via a dynamic learning rate modulation strategy. Specifically, we employ a pretrained 3D U-Net information model to quantify voxel-level prediction uncertainty, which is then used to guide the optimization of the main segmentation model. Diet-Seg is further enhanced by an RWKV-based U-Net backbone to capture global spatial dependencies and an EdgeNet module to preserve tumor boundaries through edge-aware fusion. Extensive experiments on the BraTS2018-2021 datasets demonstrate that Diet-Seg consistently outperforms state-of-the-art baselines across all tumor subregions. Notably, Diet-Seg achieves superior generalization when trained on one dataset and validated across multiple years. Moreover, the hardness maps offer interpretable insights into segmentation difficulty, potentially enabling human-AI collaboration in clinical practice. These results highlight the promise of entropy-aware training as a general strategy for robust and efficient medical image segmentation. The work is implemented in the open-source project available on GitHub (https://github.com/ManuelTurner/Diet-Seg)

Neuroanatomical segmentation in T1-weighted magnetic resonance imaging of the brain is a prerequisite for quantitative morphological measurements, as well as an essential element in general pre-processing pipelines. While recent fully automated segmentation methods based on convolutional neural networks have shown great potential, these methods nonetheless suffer from severe performance degradation when there are mismatches between training (source) and testing (target) domains (e.g. due to different scanner acquisition protocols or due to anatomical differences in the respective populations under study). This work introduces a new method for unsupervised domain adaptation which improves performance in challenging cross-domain applications without requiring any additional annotations on the target domain. Using a previously validated state-of-the-art segmentation method based on a context-augmented convolutional neural network, we first demonstrate that networks with better domain generalizability can be trained using extensive data augmentation with label-preserving transformations which mimic differences between domains. Second, we incorporate unlabelled target domain samples into training using a self-ensembling approach, demonstrating further performance gains, and further diminishing the performance gap in comparison to fully-supervised training on the target domain.

AO_SCPLOWBSTRACTC_SCPLOWMagnetic resonance imaging (MRI) is a cornerstone of modern neuroimaging, where accurate segmentation of brain structures and lesions is essential for diagnosis, treatment planning, and clinical research. However, most current foundation models are trained on mixed-organ datasets, while the anatomical structures of the brain differ substantially from those of other organs such as the lungs and kidneys. As a result, these models often struggle to adapt to the distinctive characteristics of brain tissue. In this work, we present Brain-SAM, a model tailored for brain MRI segmentation. Brain-SAM extends the Segment Anything Model 2 (SAM2) framework by enabling the Hiera encoder to directly process 3D volumetric data and introducing a UNETR-inspired decoder for hierarchical feature decoding. The model preserves the interactive segmentation paradigm of SAM while also supporting fully automatic segmentation. Trained on multiple brain MRI datasets covering brain tumors, stroke, and epilepsy, Brain-SAM demonstrated superior performance to state-of-the-art methods. Compared with nnU-Net, it achieved Dice scores improvements of 22%, 9%, and 6% on epileptic lesions, brain metastases, and meningiomas, respectively. Notably, Brain-SAM showed clear advantages in small-lesion segmentation, achieving 15%-18% higher Dice compared with other strong baseline models. We believe that Brain-SAM may offer a useful pre-trained model for downstream brain MRI analysis tasks, and could contribute to future research and clinical applications.Our code and models are available at https://github.com/DLbrainsam/Brain-SAM.