Fully supervised polyp segmentation relies on costly pixel-level annotations. Although semi- and weakly supervised methods reduce annotation requirements, they still depend on partial mask supervision. Text-supervised segmentation is a promising alternative; however, for polyps, the key challenge is to ground instance-specific phrases to the correct lesion region under cluttered backgrounds and large appearance variations. Existing approaches often rely on coarse text-image alignment, limiting precise region-level semantic correspondence. In this paper, we propose Text-Image Co-Alignment (TICoA), a text-supervised framework for polyp segmentation. TICoA leverages large language models (LLMs)-generated structured clinical descriptions as weak supervision and formulates segmentation as a fine-grained phrase-region co-alignment problem. Through contrastive learning, TICoA explicitly associates query phrases with corresponding image regions to achieve robust semantic grounding under weak supervision. Architecturally, we adopt a State-Space Model (Mamba) to efficiently model long-range dependencies with linear computational complexity. To support effective cross-modal interaction, we further design a dedicated Mamba Fusion module with a Bi-Dimension Fusion (BiDF) strategy, which progressively propagates information along spatial and channel dimensions. Experiments on polyp datasets, with additional validation on skin lesion segmentation, demonstrate that TICoA is competitive with state-of-the-art weakly supervised methods. Our code and data are available at https://github.com/silentyuchen/TICoA.

Foundation models hold promise for specialized medical imaging tasks, though their effectiveness in breast imaging remains underexplored. This study leverages BiomedCLIP as a foundation model to address challenges in model generalizations. BiomedCLIP was adapted for automated BI-RADS breast density classification using multi-modality mammographic data (synthesized 2D images, digital mammography, and digital breast tomosynthesis). Using 96,995 images, we compared single-modality (s2D only) and multi-modality training approaches, addressing class imbalance through weighted contrastive learning. Both approaches achieved similar accuracy (multi-modality: 0.74, single-modality: 0.73), with the multi-modality model offering broader applicability across different imaging modalities and higher AUC values consistently above 0.84 across BI-RADS categories. External validation on the RSNA and EMBED datasets showed strong generalization capabilities (AUC range: 0.80-0.93). GradCAM visualizations confirmed consistent and clinically relevant attention patterns, highlighting the models' interpretability and robustness. This research underscores the potential of foundation models for breast imaging applications, paving the way for future extensions for diagnostic tasks.

Background: Traumatic brain injury (TBI) is characterized by both focal and diffuse pathology. Automated cortical thickness estimation is widely used to quantify structural brain changes following TBI; however, the impact of focal pathology such as contusions on cortical thickness estimates in TBI remains unknown. Methods: We evaluated lesion-induced bias in cortical thickness under three lesion-handling strategies in 86 TRACK-TBI participants with MRI at 2 weeks and 6 months post-injury. Cortical thickness was estimated using the ANTsNetCT longitudinal pipeline with the default pipeline (No Masking), masking lesion voxels from summarization (Atlas Masking), and masking lesion voxels from cortical thickness estimation (Full Masking). Cross-sectional and longitudinal cortical thickness in unilaterally lesioned regions were compared with their contralesional homologues using linear mixed-effects models. The effectiveness of each lesion handling strategy was then evaluated using nonparametric bootstrap analyses to test whether bias was systematically present across all regions. Results: At 2 weeks post-injury, six cortical regions demonstrated significant lesion-associated bias. Collectively across all regions, bias was observed in the No Masking and Atlas-Masking approaches. This bias was significantly attenuated in the Fully Masked approach. Longitudinally, the unmasked data also showed significant lesion-related differences in cortical thickness change across multiple temporal and frontal regions, with persistent effects in the Atlas- and Full Masking approaches. Conclusions: Contusions appear to introduce cross-sectional and longitudinal bias in cortical thickness estimates, inflating cross-sectional values and potentially exaggerating atrophy longitudinally. Excluding lesion voxels from tissue probability maps attenuates cross-sectional bias, providing a baseline that improves accuracy and interpretation of neuroimaging biomarkers in TBI.

Accurate lesion detection in breast MRI is essential for early cancer diagnosis, particularly in high-risk populations. In this study, we present a classification pipeline that adapts a pretrained foundation model-Medical Slice Transformer (MST) to perform breast lesion classification using dynamic contrast-enhanced MRI (DCE-MRI). Leveraging DINOv2-based self-supervised pretraining, the MST model provides robust per-slice feature embeddings. We extract these representations and use them to train a Kolmogorov-Arnold Network (KAN) classifier, which offers a more flexible and interpretable alternative to traditional convolutional neural networks. The KAN architecture enables localized nonlinear transformations through adaptive B-spline activations, enhancing the model's ability to distinguish between benign and malignant lesions in imbalanced and heterogeneous clinical data. Our results show that the MST+KAN pipeline outperforms MST transformer based classifier, achieving improved performance metrics (AUC=0.80 +/- 0.02) while maintaining interpretability via attention-based heatmaps. This study highlights the superiority of combining foundation model embeddings with advanced classification strategies to build robust and generalizable breast MRI analysis tools.

In this work, we propose FFDP, a set of IO-aware non-GEMM fused kernels supplemented with a distributed framework for image registration at unprecedented scales. Image registration is an inverse problem fundamental to biomedical and life sciences, but algorithms have not scaled in tandem with image acquisition capabilities. Our framework complements existing model parallelism techniques proposed for large-scale transformer training by optimizing non-GEMM bottlenecks and enabling convolution-aware tensor sharding. We demonstrate unprecedented capabilities by performing multimodal registration of a 100μm ex-vivo human brain MRI volume at native resolution – an inverse problem more than 570× larger than a standard clinical datum in about a minute using only 8 A6000 GPUs. FFDP accelerates existing state-of-the-art optimization and deep learning registration pipelines by upto 6 − 7× while reducing peak memory consumption by 20 − 59%. Comparative analysis on a 250μm dataset shows that FFDP can fit upto 64× larger problems than existing SOTA on a single GPU, and highlights both the performance and efficiency gains of FFDP compared to SOTA image registration methods.

FireANTs introduced a novel Eulerian descent method for plug-and-play behavior with arbitrary optimizers adapted for diffeomorphic image registration as a test-time optimization problem, with a GPU-accelerated implementation. FireANTs uses Adam as its default optimizer for fast and more robust optimization. However, Adam requires storing state variables (i.e. momentum and squared-momentum estimates), each of which can consume significant memory, prohibiting its use for significantly large images. In this work, we propose a modified Levenberg-Marquardt (LM) optimizer that requires only a single scalar damping parameter as optimizer state, that is adaptively tuned using a trust region approach. The resulting optimizer reduces memory by up to 24.6

We present NEUSEG, a fully automated, interpretable, and unsupervised pipeline for gray/white matter (GM/WM) segmentation in brain histopathology whole-slide images (WSIs). GM/WM segmentation in WSIs is challenging due to their gigapixel-scale resolution and substantial variability in staining, regional morphology, and underlying pathology. To address this, NEUSEG combines nuclei morphometrics with a two-component Gaussian Mixture Model and refines boundaries using a Conditional Random Field and morphological post-processing. In comparison with a supervised CNN baseline, NEUSEG achieved comparable accuracy on in-distribution slides and superior robustness under out-of-distribution conditions. Across 252 WSIs with expert-annotated regions of interest (ROIs), median annotation-to-contour distances were 32.89 μm for the GMBG boundary and 129.79 μm for the GM-WM boundary, with no significant difference across stains or pathology types and only minor regional variation. Percent area occupied between expertand segmentation-derived ROIs showed near-perfect agreement (r=0.987). NEUSEG is lightweight, CPU-operable, and processes each WSI within minutes, enabling scalable and robust GM/WM segmentation across heterogeneous histopathology datasets.

Vision foundation models are increasingly moving beyond 2D to volumetric domains such as 3D medical imaging, where unified pretraining across different imaging modalities (i.e. CT, MRI, and PET) could provide foundational models for diverse clinical tasks. However, training such models requires mixing heterogeneous imaging domains, and current mixture strategies remain largely heuristic. In this work, we observe that different medical imaging domains scale at variable rates during pretraining, and knowledge transfer between domains is strongly asymmetric: training on one domain can substantially improve another, but the reverse may be much weaker. Interestingly, both MAE reconstruction loss and cross-domain transfer follow predictable power-law trends with domain-specific behaviors. Motivated by these findings, we formulate data allocation as a scaling-law optimization problem. The derived allocations reveal an interpretable hub-and-island structure: highly transferable domains emerge as hubs that benefit many others and deserve strategic allocation, while isolated domains act as islands requiring direct investment. Empirically, transfer-aware allocation outperforms data-proportional sampling by up to 58

In modeling complex probability distributions, normalizing flows provide exact-likelihood, bijective mappings between empirical data and tractable latent spaces. Building on this foundation, latent-aligned multiview normalizing (LAMNr) flows leverage these salient properties to learn shared latent subspaces across heterogeneous, multimodal datasets while simultaneously topologically unfolding the sampled data manifold into a continuous vector space. Formal latent-alignment constraints are used to model shared structural features separate from view-specific variations, coordinating latent projections into a shared geometric subspace. By applying this transformation in the context of biological imaging, the framework establishes a potential basis for a deep learning interpretation of foundational computational anatomy concepts, such as the population template, latent distances, and geodesic pairwise image interpolation. Additionally, the proposed framework enables closed-form conditional modeling for exact cross-view imputation and other latent space manipulations. Evaluations and illustrations on both imaging-derived phenotypes (IDPs) and multimodal MRI demonstrate the proposed framework and potential applications. To further motivate our work, we provide a robust and comprehensive, 2D and 3D open-source implementation in PyTorch, natively integrated with the ANTsX ecosystem (i.e., ANTsTorch) for efficient training and subsequent data transformation, manipulation, and analysis.

Accurate segmentation of low-contrast lesion boundaries in medical images remains a key challenge for learning-based segmentation methods. Existing approaches often lack sufficient structural modeling capabilities when dealing with lesions that exhibit blurred boundaries and irregular shapes, making it difficult to accurately recover complex anatomical structures. The level set method, known for its strong adaptability to topological variations and effective modeling of continuous boundaries, has been widely used in the segmentation of structurally complex medical images. However, traditional level set algorithms suffer from high inference complexity and strong dependence on initialization, which limits their practical applicability. To address these limitations, this paper proposes a diffusion-guided level set evolution framework (LSDiff) for structural modeling and segmentation of low-contrast lesion boundaries. The proposed method leverages the intrinsic consistency between diffusion models and level set methods in terms of iterative mechanisms and partial differential equation (PDE) structures by embedding the level set energy evolution process into the sampling trajectory of the diffusion model. A joint energy functional is constructed to guide the level set surface to evolve progressively from a randomly initialized state toward a clear and structurally consistent boundary, enabling accurate fitting of complex lesion contours. Extensive experiments are conducted on four representative medical image datasets—ISIC 2016, Kvasir-SEG, BUSI and BraTS 2020. The results demonstrate that the proposed LSDiff framework significantly outperforms a range of mainstream segmentation methods in terms of boundary accuracy and structural integrity, validating its effectiveness for low-contrast lesion boundary segmentation.

Globular glial tauopathy is a 4-repeat tauopathy associated with heterogenous clinical syndromes, including primary progressive aphasia. Iron-reactive gliosis in mid-to-deep cortical layers has previously been reported in this disorder, but detailed anatomic localization and its relationship to clinical symptoms is understudied, particularly within the anatomic framework of primary progressive aphasia. In a series of five autopsy-confirmed patients with globular glial tauopathy and one healthy control, we utilize ultra-high-resolution whole-hemisphere ex vivo 7 Telsa MRI and digital pathology to study whole-hemisphere and local laminar/cellular patterns of pathology within affected cortex. We find signature laminar patterns of iron-rich gliosis localized to brain regions implicated in distinct clinical aphasia syndromes between patients: patients who presented with non-fluent aphasia had iron-rich gliosis pathology localized to inferior and superior frontal and motor regions while iron-rich pathology was largely localized to the anterior temporal lobe in a patient with the semantic variant. Moreover, in one patient with non-fluent aphasia and additional iron-sensitive 7 Telsa MRI during life, we find evidence of antemortem iron-rich pathology in the same frontal regions observed post-mortem. These data suggest that focal neuroinflammation and iron dysregulation may contribute to the clinical expression of tauopathies and be detectable during life to improve diagnosis.

Low bone density (LBD), defined as the combined group of osteoporosis and osteopenia, remains substantially underdiagnosed despite the presence of effective screening and treatment options. Our study evaluates whether fully automated artificial intelligence-based segmentation of L1 vertebral body attenuation on computed tomography (CT) enables scalable opportunistic LBD screening across two large academic institutions. Overlap between CT-defined LBD and ICD-coded diagnoses was limited, with 59–76% of imaging-defined cases lacking a corresponding clinical diagnosis, indicating a large undetected population. Phenome-wide association studies in both cohorts linked LBD to osteoporosis, respiratory failure, sepsis, coagulation defects, and cardiovascular disease, while genome-wide analysis identified no significant loci. Additionally, median L1 vertebral attenuation declined most sharply between the fifth and sixth decades, indicative of a potentially earlier decline of bone density than the start of current gold standard osteoporosis screening. Combined, these findings support automated opportunistic CT analysis as a feasible, scalable strategy for earlier LBD detection and for revealing clinically relevant population-level associations.

Abstract Image matching is a fundamental task in quantitative biomedical and biological image analyses, enabling researchers to compare, integrate, and interpret imaging data across subjects, time points, modalities, and experimental conditions. Existing state-of-the-art registration methods are slow due to inefficient implementations and poor convergence rates because of the ill-conditioned nature of the optimization problem. Deep learning methods offer fast inference but require extensive training time, substantial inference memory, and fail to generalize across long-tailed distributions or diverse image modalities, necessitating costly retraining. We address these challenges by proposing FireANTs, a training-free, GPU-accelerated, multi-scale adaptive Riemannian optimization algorithm for fast and accurate dense diffeomorphic image matching. FireANTs more than doubles the speed of the community standard ANTs registration tool on a CPU, and is two orders of magnitude faster on a GPU. On the GPU, FireANTs performs competitively with deep learning methods on inference runtime while consuming up to 10 × less memory. FireANTs demonstrates robustness on a wide variety of matching problems across modalities, species, and organs, without any domain-specific training or tuning. Our framework allows hyperparameter grid search studies with less resources and time compared to traditional and deep learning registration algorithms alike.

Frontotemporal lobar degeneration (FTLD) is defined neuropathologically by misfolded tau (FTLD-tau) or TAR DNA-binding protein of 43 kDa (FTLD-TDP). However, we lack biomarkers that can distinguish them in vivo which is a major barrier to effective disease-modifying treatment trials. Based on neuropathological evidence of distinct patterns of cellular degeneration, more prominent in white matter (WM) for FTLD-tau relative to FTLD-TDP, we hypothesized that diffusion MRI (dMRI) measures of white matter microstructure would help dissociate FTLD-tau and FTLD-TDP during life. We evaluated dMRI for 157 individuals with FTLD-Tau (n=66, autopsy-confirmed sporadic=51, MAPT=15) and FTLD-TDP (n=91, autopsy-confirmed sporadic=25, GRN=19, C9orf72=47) defined by a pathogenic genetic mutation or autopsy confirmation. Images were preprocessed using QSIprep and reconstructed using Generalized Q-sampling Imaging. Tensor-based scalar maps of fractional anisotropy (FA) and mean diffusivity (MD) were then calculated. Deterministic tractography was run and white matter bundles of pathogenic interest based on literature were segmented with DSIstudio using template-based tracking. Overall, FTLD-Tau individuals showed worse white matter integrity compared to those with FTLD-TDP with lower FA in the superior longitudinal fasciculus (SLF), uncinate fasciculus and fronto-parietal cingulum; and higher MD in the uncinate fasciculus. Even within sporadic autopsy-confirmed individuals, FTLD-Tau subtypes had significantly lower FA in the SLF, fronto-parietal cingulum, and left corticostriatal tract relative to FTLD-TDP. Our results demonstrate that white matter microstructure can distinguish between FTLD-Tau and FTLD-TDP pathologies in a tract-specific manner with greater evidence of white matter degeneration in disease-associated regions in FTLD-Tau. Future work examining longitudinal associations with disease progression and white matter microstructure could aid in non-invasive differentiation of these subtypes enabling more sensitive treatment monitoring.

BACKGROUND:Deterioration of right ventricular (RV) function in repaired tetralogy of Fallot (rToF) is poorly understood. Cardiovascular magnetic resonance (CMR) is used for monitoring, but its analysis is user-dependent and time-consuming. We sought to automate the analysis of CMR using machine learning and to identify imaging traits associated with adverse RV remodeling in the natural history of rToF. METHODS:A longitudinal cohort of rToF patients underwent CMR at the Children's Hospital of Philadelphia. The nnU-Net method was used to train a machine learning model to segment the left ventricular (LV) blood pool, LV myocardium, and RV blood pool from two-dimensional short-axis CMR images. Conventional and novel measures were calculated and studied in association with remodeling rates using multivariable linear regression. Remodeling rates were calculated as ((Variablescan2 - Variablescan1)/years between scans) for the variables end-diastolic volume index (EDVi), end-systolic volume index (ESVi), stroke volume index (SVi), ejection fraction (EF), and peak systolic dV/dt. RESULTS:The cohort was comprised of 758 patients, of whom 152 had 2 analyzable scans. Thirty-six patients underwent pulmonic valve replacement (PVR) between scans. Compared to patients with no intervention (representing the natural history of rToF), patients with PVR had significantly lower remodeling rates for RVEDVi, RVESVi, RVSVi, and absolute peak systolic RV dV/dt, while RVEF and left-sided metrics did not differ between groups. In 116 patients without PVR between scans, RV remodeling rates were negatively associated with baseline LV mass index, LVEDVi, LVSVi, and absolute peak systolic LV dV/dt. CONCLUSION:We demonstrated that rToF patients with two CMR scans and PVR have significant differences in and opposite directions of RV remodeling rates compared to those with no intervention. We also showed that several left-sided measures of structure and function were associated with RV remodeling rates, indicating the importance of baseline LV measurements in characterizing future risk of adverse RV remodeling.

Paradoxical reductions in brain volume in individuals with Alzheimer’s disease treated with anti-amyloid therapy have been reported in clinical trials. It has been hypothesized that these changes could partly be attributed to processes related to the clearance of amyloid plaques. Further, the presence of amyloid-related imaging abnormalities (ARIA) has been associated with greater ventricular expansion in clinical trials of anti-amyloid immunotherapy, and may also be an important factor driving brain volume loss. Here we measure longitudinal changes from structural MRI during the course of anti-amyloid treatment in brain regions known to accumulate amyloid plaques and ask whether these regions show differences in observed changes depending on presence of ARIA. We analyzed 125 MRI images from 45 patients undergoing surveillance scanning during the course of lecanemab therapy. Average duration between the earliest and latest scans was 181 days. For each patient, images were included if from an identical MRI protocol and scanner. Measures of gray matter thickness in 100 brain regions were obtained using an unbiased longitudinal pipeline. Regions were nominally dichotomized into amyloid-accumulating and non-amyloid-accumulating based on composite regions used for determining amyloid positivity from PET imaging. Thickness was analyzed using linear mixed effects model that included time and presence of ARIA (ARIA-E or ARIA-H) as explanatory variables and age and sex as nuisance covariates. Both amyloid-accumulating and non-amyloid-accumulating regions showed thinning during the treatment period. However, there was a significant interaction between time and ARIA status only in amyloid-accumulating regions such that patients with ARIA showed greater observed thinning. Hippocampal volume also showed a decrease over time, but there was no effect of presence of ARIA. These data show that presence of ARIA may play a role in observed brain parenchymal changes during anti-amyloid treatment. The specificity of the effects in amyloid-accumulating brain regions point to an interaction between amyloid clearance and the emergence of ARIA as confounding factors in measuring longitudinal change. While the underlying mechanism remains unclear, better characterization of such effects can help disentangle their influence on MRI-based measures of change and ultimately improve the usability of MRI as a marker for tracking potential treatment effects.

BACKGROUND:The organization of the brain into distinct networks increases (i.e., differentiation) during development and decreases (i.e., de-differentiation) during healthy aging, changes that are associated with improvements and worsening of cognition, respectively. Given that behavioral variant frontotemporal degeneration (bvFTD) is a neurodegenerative disease associated with executive dysfunction and selective vulnerability of the salience network, we tested the hypotheses that bvFTD structural networks are de-differentiated compared to cognitively normal controls (CNC) and that network de-differentiation relates to worse executive function. METHODS:In a sample of 90 patients with bvFTD and 71 age-matched CNC with diffusion MRI data we generated probabilistic tractography maps and calculated system segregation, a metric that compares within-network to between-network connectivity, to reflect the extent to which brain networks were differentiated. Patients with bvFTD also completed tests of executive function (digit span backwards, phonemic fluency, category fluency) and a control task (lexical retrieval). We assessed group differences in system segregation, reflecting network differentiation, and, within bvFTD, associations between system segregation and neuropsychological test performance. RESULTS:Compared to CNC, patients with bvFTD exhibited lower system segregation of the salience (p < 0.001) and global brain network (p = 0.008). In bvFTD, lower salience network system segregation was associated with worse executive function (pcorrected = 0.021) but not lexical retrieval. CONCLUSIONS:Results demonstrate associations between executive dysfunction and salience network de-differentiation in patients with bvFTD. Our findings indicate that brain network de-differentiation, reflecting reduced neural capacity for specialized processing, may contribute to the emergence of executive dysfunction in bvFTD.

Large-scale efforts by the BRAIN Initiative Cell Census Network (BICCN) are generating a comprehensive reference atlas of cell types in the mouse brain. A key challenge in this effort is mapping diverse datasets, acquired with varied imaging, tissue processing, and profiling methods, into shared coordinate frameworks. Here, we present mouse brain mapping pipelines developed using the Advanced Normalization Tools Ecosystem (ANTsX) to align MERFISH spatial transcriptomics and high-resolution fMOST morphology data to the Allen Common Coordinate Framework (CCFv3), and developmental MRI and LSFM data to the Developmental CCF (DevCCF). Simultaneously, we introduce two novel methods: 1) a velocity field-based approach for continuous interpolation across developmental timepoints, and 2) a deep learning framework for automated brain parcellation using minimally annotated and publicly available data. All workflows are open-source and reproducible. We also provide general guidance for selecting appropriate strategies across modalities, enabling researchers to adapt these tools to new data.

BACKGROUND:Thoracic aortic dissection is a life-threatening condition that often occurs in the presence of aortic dilation. However, currently there are limited clinical risk factors beyond aortic diameter (AoD) used to determine individual-level dissection risk. OBJECTIVES:The purpose of this study was to determine whether common variant genetics can be used to improve identification of individuals most at risk for dissection. METHODS:A genome-wide association study (GWAS)-by-subtraction was performed to characterize the diameter-independent genetics of thoracic aortic dissection by subtracting a GWAS of AoD from a GWAS of thoracic aortic aneurysm and dissection. A polygenic risk score (PRS) was calculated using the PRS-Continuous Shrinkage statistical package and applied to Penn Medicine BioBank participants. Statistical analysis was performed in R version 4.3.2. RESULTS:We identified 43 genetic risk loci associated with dissection and derived a "Dissection-PRS" from our GWAS-by-subtraction. In the Penn Medicine BioBank, the Dissection-PRS associated with prevalent dissection (OR: 2.13 per 1 SD increase in Dissection-PRS; 95% CI: 1.91-2.39; P < 0.001). When adjusting for risk factors including AoD, the association of the Dissection-PRS with prevalent dissection was attenuated but remained statistically robust (OR: 1.62 per 1 SD increase in PRS; 95% CI: 1.36-1.94; P < 0.001). The addition of the PRS to a model containing age, sex, and clinical risk factors substantially improved model discrimination (base model area under the receiver operator characteristic curve = 0.676; 95% CI: 0.651-0.702; with addition of PRS area under the receiver operator characteristic curve = 0.723; 95% CI: 0.702-0.744). CONCLUSIONS:A common-variant PRS can predict aortic dissection in a diverse population.