Medical Image Analysis

Accurate prenatal screening from ultrasound imaging is severely constrained by the extreme spatial heterogeneity of fetal structures and the data scarcity inherent to rare congenital anomalies. To overcome these computational bottlenecks, we present RadarMAE, a data-efficient self-supervised representation learning framework tailored for fetal ultrasound analysis. By employing a novel, structure-aware masking strategy, RadarMAE enables robust pre-training on highly limited clinical datasets. Benchmarking the framework on first-trimester screening for trisomy 21 (Down syndrome), our model achieved exceptional diagnostic performance on an internal cohort, with an area under the curve (AUC) of 0.99, 97% accuracy, 94% sensitivity, and 98% specificity. Crucially, through domain-adversarial adaptation, the framework demonstrates robust cross-institutional generalization under extreme source-target data imbalance, bounding performance degradation to less than or equal to 0.04 across all key metrics in external testing. Furthermore, by integrating spatial registration with explainable AI, we constructed the first population-level facial risk atlas for trisomy 21. This data-driven interpretability framework validates standard clinical markers while revealing the comparable diagnostic weight of underutilized features, such as mandibular dysmorphology. Ultimately, this approach provides a scalable, transparent, and high-performance solution for equitable prenatal diagnostics and phenotypic biomarker discovery.

Timely detection of fetal distress during labor is a central pursuit of modern obstetric care. The primary measure used in this regard is fetal heart rate and contraction monitoring (echocardiotocography or CTG), which provides a continuous signal during labor. The vast majority of machine learning applications to CTG assume access to a 30 to 60 minute window just prior to delivery, which makes an untenable assumption that the timing of delivery is known. In this work, we reframe the problem as a time-to-event prediction task. We develop deep time-to-event models to predict the joint probability of imminent delivery and fetal acidosis, a biomarker of fetal hypoxia collected from umbilical cord gas after delivery. We develop a framework for this problem, drawing on deep discrete-time survival methods and marked point processes, adapting each to this particular setting (i.e., continuous physiological signals, fully observed events, partially observed labels). One approach dubbed Marked DeepHit accurately estimates delivery within one hour at multiple elapsed-time landmarks, with test-set AUROC of 0.837 [95% CI 0.692--0.970] and 0.935 [0.858--0.964] at 6 and 12 hours into labor, respectively. We find that these models also achieve competitive performance near delivery time compared to prior models trained specifically on the end of tracing, and outperform other clinical feature-based proposals, both on internal and external validation at a second site. These results give a proof-of-principle to support future prospective deployment of AI-EFM during labor.

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.

Cardiac imaging enables quantitative assessment of cardiac structure and function but remains constrained by cost, infrastructure and specialist expertise. In contrast, electrocardiogram (ECG) is widely accessible yet underexploited, despite encoding latent information about cardiac physiology. Here we introduce visionECG, a conditional flow matching framework that learns a probabilistic mapping between two biological distributions - the space of cardiac electrical signals and the space of cardiac geometries. Using 71,132 paired ECG and cardiac mesh sequence datasets from the UK Biobank, with external assessment in 5,000 patients with ECG-echocardiogram pairs, the model reconstructs quantitatively accurate spatiotemporal representations of the left ventricle using ECG inputs and basic demographic information alone. These reconstructions enable discrimination of structural abnormalities and disease labels, provide visualisations of functional abnormalities, and support flexible quantification of both global and regional parameters. By reframing the ECG as a generative source of patient-specific left ventricular geometry and motion, this work establishes a scalable framework for translating low-dimensional signals into high-dimensional, physiologically grounded structured representations.

Text-driven medical image segmentation demands unified models based on multi-modal learning, capable of processing diverse imaging modalities while providing spatially reliable prediction estimates for clinical practice. Nevertheless, existing vision-language model solutions suffer from poor domain generalization and exhibit limited capability for uncertainty estimation to model the cumulative unreliability of cross-modal alignment across network layers. In this paper, we propose UniMedSeg, a unified vision-language framework for multi-modal medical image segmentation built upon frozen pre-trained vision-language backbones. Our framework incorporates modality-aware adaptation, cross-layer uncertainty-guided encoding, and a frequency-spatial dual-prompt decoder, wherein the progressive uncertainty propagation mechanism of the encoder suppresses misleading cross-modal features, whereas the decoder disentangles anatomical structure and boundary details in the frequency domain under dual textual prompts. Auxiliary losses further sharpen boundary representations and calibrate pixel-wise uncertainty maps. Extensive experiments demonstrate that UniMedSeg preserves competitive in-domain segmentation performance, improves domain generalization, and yields clinically useful spatial risk estimation.

Purpose: To develop and evaluate a recurrent neural network that operates directly on undersampled golden-angle radial k-space, enabling low-latency cardiac cine reconstruction for interventional cardiac MRI (iCMR). Methods: KOSTAL (K-space Optimized Spatial Temporal Architecture for Low latency) combines a per-coil k-space convolutional gated recurrent unit (convGRU), which learns a mask-conditioned view-sharing rule over the accumulated k-space, with an image-space recurrent module. It was trained and tested on the open-access multi-coil OCMR dataset using 10 simulated golden-angle spokes per frame against a 1000-spoke reference. Ablations varied recurrent module architecture and placement; temporal behavior was assessed using a imaging-plane transition test. Generalization was evaluated without retraining in 23 pediatric patients prospectively scanned during free-breathing using a radial golden-angle bSSFP sequence with binned XD-GRASP and clinical breath-held cine serving as references. Results: At 10 spokes per frame, KOSTAL reached a median SSIM of 0.907 (gridding: 0.622) in under 34 ms of processing time per frame. Image quality (IQ) matched gridded results formed from 7-11 times more spokes. Recurrence in k-space rather than image space accounted for most of this gain (0.891 versus [≤]0.650). IQ recovered from an abrupt plane transition in 75 ms, versus 303 ms for 200-spoke gridding. KOSTAL had the best agreement with the binned reference and the highest long-axis CNR. KOSTAL accurately tracked left-ventricular blood-pool volume (end-diastolic volume bias -0.7 mL) with a shorter lag (71 ms) than iterative reconstruction (239-287 ms). Conclusion: Learned, mask-conditioned recurrence in k-space yields high image quality at a latency compatible with interventional guidance.

Automated segmentation of Optical Coherence Tomography (OCT) images is a critical component of structural biomarker extraction for retinal diagnostics. Deep learning models achieve state-of-the-art performance on controlled datasets, yet exhibit unpredictable failures on real-world data. Current quality gates rely on device-reported scan quality scores, which have been shown to be unreliable predictors of segmentation performance. We define scan quality in a task-specific sense, that is, whether a given B-scan will yield a reliable segmentation from a particular trained model. Under this definition, we systematically evaluate No-Reference Image Quality Assessment (NR-IQA) metrics, general-purpose pretrained representations, and domain-specific pretrained representations as alternative quality gates. We use choroid segmentation as the prototype task, with a dataset of 6,076 OCT B-scans from 80 subjects. These quality gate candidates are evaluated at three levels: scalar metrics (BRISQUE, NIQE, PIQE, SNR, PSNR), supervised linear probing, and unsupervised partitioning (K-Means) of feature vectors and learned representations. All scalar NR-IQA metrics proved inadequate (|r| < 0.20). General-purpose pretrained representations (EfficientNet-B0, ResNet-50, ViT-B/16) outperform NR-IQA, achieving ROC-AUC up to 0.77, indicating that learned representations are better suited to task-specific quality gating than hand-crafted scalar statistics. Retinal foundation models (FMs) further improve performance: RETFound, an OCT-specific FM, achieves ROC-AUC of approximately 0.81. Unsupervised K-Means partitioning of the embeddings recovers the same hierarchy geometrically: only the two retinal foundation models produce quality-aligned clusters that exceed a patient-level permutation null, while NR-IQA and general-purpose pretrained feature spaces do not, suggesting that domain-specific pretraining provides an additional benefit beyond general-purpose learned representations.

Coarse to fine segmentation is a natural strategy for three dimensional lesion delineation becausefull resolution processing of an entire scan is expensive while only a small fraction of the volume contains pathology. Its main failure mode, however, is structural: a lesion missed by the coarse stage is usually absent from the region of interest and therefore cannot be recovered by the fine stage. This problem is most severe for small, low contrast, and multifocal lesions. We introduce RECALLREFINE, a coverage constrained coarse to fine framework that treats candidate-region selection as a recall critical learning problem rather than a deterministic crop around the coarse mask.A full-volume network first predicts coarse lesion probabilities and multi-scale features. A candidate score network then combines coarse probability, predictive uncertainty, feature novelty, and spatial context to score overlapping 3D blocks. During training, a differentiable selector is optimized with a component-wise coverage objective that requires every annotated lesion instance to remain covered by at least one selected block under a fixed refinement budget. Selected blocks are processed at native resolution by a local refiner with a context-ring branch, and overlapping predictions are merged back into the full volume. Across four public 3D lesion benchmarks, RECALLREFINE improves mean tumor Dice from 76.5% for the strongest matched coarse-to-fine baseline to 78.4%, while increasing sensitivity for lesions below 10 mm from 60.9% to 70.8%.At a 20% candidate budget, the selector covers 94.4% of lesion instances, substantially reducing irreversible coarse-stage misses. The results suggest that the central problem in cascaded lesion segmentation is not only how to refine a detected lesion, but how to guarantee that difficult lesions are still represented in the refinement set.

The detection of small and subtle lesions in high-resolution 3D volumes is a highly relevant, yet far from solved task in biomedical imaging. We here address a specific task in detecting certain types of epileptogenic lesions through our novel semi-supervised spatially context aware transformer (SpyCAT) approach to anomaly detection. SpyCAT is modeling-based in the sense that it builds on specific assumptions that constitute what is normal and what constitues relevant deviations from normality. We explicitly use these assumptions to justify the inductive bias of our anomaly detection approach. The resulting SpyCAT system is patch-based and uses a transformer architecture to process discrete tokens obtained from a vector quantizing variational autoencoder, which produces counterfactual patches through full 3D convolutions of each patch. We evaluate our approach on the grounds of point-annotations of two subtypes of epileptogenic lesions, using validation measures that build on the Metrics Reloaded framework, showing that SpyCAT can reliably identify and localize the lesion types under consideration, and outperforms state-of-the-art reference methods.

BACKGROUND: Coronary angiography remains the reference standard for diagnosing coronary artery disease and guiding revascularization, yet its interpretation requires expert integration of multi-view anatomy, lesion morphology and procedural context. Existing artificial intelligence approaches are largely task-specific, annotation-dependent and limited in capturing the semantic relationship between angiographic findings and interventional decision-making. Whether large-scale vision-language pretraining can enable transferable foundation-model representations for invasive coronary imaging remains unknown. METHODS We developed CAG-MIND, a domain-specific vision-language foundation model for coronary angiography, using 135,475 CAG examinations paired with procedural reports, comprising 812,850 angiographic videos from Zhongshan Hospital and Shanghai Geriatric Medical Center. Each case consisted of standardized six-view angiographic acquisitions paired with structured procedural semantics extracted from routine reports using a large language model-assisted pipeline. The model was pretrained by aligning multi-view angiographic representations with report-derived semantic embeddings through bidirectional contrastive learning. Performance was evaluated under zero-shot and supervised fine-tuning settings across 11 downstream tasks grouped into structural abnormality detection, atherosclerotic plaque assessment, and interventional decision prediction, using both an internal validation cohort and an independent external test cohort. RESULTS CAG-MIND demonstrated robust performance across all three task categories. In the zero-shot setting, the model achieved mean AUROCs of 0.686 in the internal validation cohort and 0.745 in the external test cohort, indicating transferable multimodal representations without task-specific supervision. Following supervised fine-tuning, the mean AUROC increased to 0.827 and 0.846, respectively, with excellent performance for coronary stenosis detection (AUROC 0.940 in both cohorts), balloon/stent prediction (0.900 and 0.907), and CABG recommendation (0.877 and 0.875). Compared with representative biomedical vision-language models and conventional image-based architectures, CAG-MIND consistently achieved superior performance in both zero-shot and supervised settings and remained superior to fully fine-tuned competing models when trained with only 10% of the labelled data. Grad-CAM visualization demonstrated anatomically plausible lesion-focused attention, supporting the interpretability of the learned representations. CONCLUSIONS CAG-MIND is, to our knowledge, the first large-scale vision-language foundation model for coronary angiography trained at more than 100,000-patient scale. By aligning standardized multi-view angiographic videos with report-derived procedural semantics, CAG-MIND enables robust zero-shot transfer, data-efficient fine-tuning and cross-center generalization. These findings support domain-aligned multimodal pretraining as a scalable foundation-model paradigm for invasive cardiovascular imaging and future cath-lab decision support.

Brain foundation models have shown promising advances in learning generalizable representations from functional magnetic resonance imaging (fMRI) data. However, existing models are typically tied to specific brain atlases and cannot be applied to fMRI data represented using different atlases without full retraining, thereby severely limiting their broader utility across neuroimaging studies. To address this bottleneck, we introduce AREION, atlas-agnostic representation learning of brain connectomes. AREION overcomes atlas dependency through two integrated innovations. First, a Brain Region Contextualizing technique maps brain regions and connectivity entries from diverse atlases into a unified representation space based on their spatial, functional, and morphological characteristics. Second, a Hierarchical Brain Transformer Encoder processes region-specific connectivity profiles as variable-length token sequences and jointly captures local functional structure and global brain functional organization, conditioned on brain regional context, without assuming a fixed number of regions. Together, these two techniques enable the learning of generalizable representations across brain atlases with different parcellation resolutions and connectivity dimensions. AREION is pretrained on a large-scale multi-atlas cohort of fMRI data via self-supervised modeling and evaluated across diverse downstream tasks. Extensive experiments show that it outperforms strong supervised baselines and fMRI foundation models while using substantially fewer parameters, demonstrating strong transferability to atlases unseen during pretraining, generalization to external datasets, and excellent interpretability at the brain regional level.

Objective: We developed a data-efficient deep learning framework for three-dimensional segmentation of pathological musculoskeletal anatomy from magnetic resonance imaging (MRI) when only limited manual annotations are available. Methods: We developed a Physics-Informed Latent-Regularized U-Net (PILR-U-Net) that combines transfer learning from healthy MRI, latent-space anatomical regularization, and elasticity-based physics-informed constraints. The physics-informed loss enforces mechanical equilibrium, near-incompressibility, and spatial smoothness of predicted deformation fields. The framework was evaluated on MRI datasets from 50 participants with cerebral palsy across 15 lower-limb musculoskeletal structures using sparse manual annotations. Performance was assessed using volumetric overlap, boundary accuracy, volume error, sensitivity, and precision. Ablation experiments evaluated the individual contributions of latent-space and physics-informed regularization. Results: Our experimental results show accurate segmentations with three-dimensional Dice coefficients ranging from 0.750 to 0.943 across evaluated structures, while most structures exhibited low surface-distance errors. Predicted deformation fields maintained positive Jacobian determinants near unity and smooth strain-energy distributions. Ablation analysis showed that both regularization components improved performance, with removal of physics-informed regularization producing the largest reductions in Dice accuracy and increases in boundary error. Conclusion: PILR-U-Net enables accurate and anatomically plausible segmentation of pathological musculoskeletal MRI under sparse supervision. Significance: Incorporating anatomical priors and biomechanical constraints into deep segmentation networks may reduce dependence on extensive pathological annotations and support patient-specific musculoskeletal modeling and clinical analysis.

Early and accurate risk stratification of cardiovascular disease (CVD) is crucial to initiate timely preventive interventions. As large-scale multimodal clinical cohorts become increasingly available, there is growing interest in whether incorporating additional sources of information can improve CVD risk stratification. Cine cardiac MR (CMR) represents a compelling example of such a source, as it captures objective, high-dimensional structural and functional information about the heart, independent of patient-reported data. In this study, we deploy a flexible vision-tabular method to incorporate cine CMR into CVD risk assessment together with structured clinical data. Using a large prospective imaging cohort from the UK Biobank, we show that cine CMR encodes CVD risk beyond established risk scores, increasing AUROC by 0.036 over SCORE2, the best-performing traditional risk score (0.742 vs. 0.706, \textit{p} = 0.04). Furthermore, we find that cine CMR achieves risk discrimination capabilities on par with automated, image-derived phenotypes, removing the dependency on segmentation pipelines. Lastly, we demonstrate that integrating cine CMR with clinical variables through a vision-tabular learning framework stabilizes risk prediction under real-world conditions of incomplete tabular data, a common challenge in clinical practice. Together, these findings position cine CMR as a promising modality for CVD risk assessment.

Automated detection of brain tumors from Magnetic Resonance Imaging (MRI) can accelerate diagnosis and reduce inter-reader variability, yet many existing studies report only top-line accuracy on small datasets, omit efficiency analysis, and provide no interpretability, limiting their clinical credibility. We present a reproducible, comparative, and explainable transfer- learning framework for binary brain-tumor classification. Our framework (i) standardizes a configurable preprocessing pipeline combining CLAHE contrast enhancement and unsharp-mask sharpening, (ii) evaluates a custom CNN baseline and pretrained backbones under an identical training budget, (iii) reports a full metric suite (accuracy, precision, recall, F1, ROC-AUC, PR-AUC, parameter count, and inference latency), and (iv) applies Grad- CAM for spatial interpretability. On a public 253-image MRI dataset (38-image held-out test set), MobileNetV2 achieves the best overall performance (94.74% accuracy, 0.994 ROC-AUC, 0.996 PR-AUC) with only 2.59M parameters and 5.9 ms per- image inference, making it the most deployment-friendly model. Larger backbones (Xception, EfficientNetB0) and the custom CNN converge to degenerate all-positive predictions under the same limited budget, illustrating the small-data overfitting risk that accuracy-only reporting conceals. Grad-CAM confirms that the best model attends to the tumor region. All source code, con- figuration files, and trained evaluation scripts are publicly avail- able at https://github.com/blck-iris/explainable-brain-tumor-mr

Magnetic resonance imaging (MRI) exhibits substantial scanner and protocol dependent appearance variations, making harmonization challenging without distorting patient specific anatomy, particularly in abdominal imaging. We propose AC-Diff, an anatomy--contrast disentangled diffusion framework that formulates MRI harmonization as a factor-specific generative intervention. Using aligned multi-contrast supervision and cross-patient factor swapping, AC-Diff learns to separate anatomical structure from acquisition-dependent contrast. Unlike conventional image or latent diffusion, AC-Diff restricts stochastic generation to the contrast subspace while the source anatomy bypasses diffusion and is directly reused during reconstruction. Experiments on an in-house multi-vendor cohort and the external Duke Liver MRI dataset demonstrate improved target-domain alignment with strong structural preservation. In downstream liver segmentation, AC-Diff improves Dice from 0.942 for the original inputs to 0.954 for the harmonized images. These results support contrast-specific latent generation as a promising approach to anatomy-preserving MRI harmonization.

Standard cardiac biomarkers such as ejection fraction, ventricular volumes, and strain compress complex three-dimensional heart dynamics into a small set of predefined measurements, leaving clinically important heterogeneity unrecognised. Here we propose a personalised normative cardiac framework in which each individual is compared to a digital twin representing expected healthy cardiac function given their age, sex, and body habitus. We show that structured patterns of deviation from these normative twins, discovered through systematic exploration of a generative cardiac dynamics space, constitute novel dynamic phenotypes that are not captured by conventional biomarkers. Applied independently to two UK Biobank cardiac cohorts with distinct shape representation, a 128-dimensional four-chamber statistical shape descriptor (HeartSSM, n = 52,365) and a 64-dimensional deep generative latent space from a conditional variational mesh autoencoder (MeshHeart cVAE, n = up to 39,376), we identify reproducible, continuous deviation-from-normative-twin scores (three discrete deviation phenotypes in the HeartSSM cohort) that improve cardiovascular risk stratification beyond ejection fraction, ventricular volumes, myocardial mass, and strain. The strongest effect was seen for incident atrial fibrillation among individuals with preserved ejection fraction (at least 50%): incremental C-index delta C = +0.068 (5-fold cross-validated delta C = +0.063; likelihood-ratio P = 3.3 x 10^-8), with a separate, smaller effect for incident stroke in the same preserved-EF subgroup in the independent MeshHeart cVAE representation (which showed no signal for atrial fibrillation) (delta C = +0.024, P = 2.6 x 10^-4). Both effects were checked against placebo controls in which the deviation scores were replaced by Gaussian noise. To test whether the normative twin generalises beyond UK Biobank's specific scanning and processing pipeline, we applied the MeshHeart cVAE normative twin, unmodified, to an independent external cohort of healthy volunteers imaged at a different hospital (GenScan/Hammersmith, n = 315): deviation scores in this cohort were of comparable magnitude to the UK Biobank healthy reference (ratio of means = 1.09; Mann-Whitney P = 1.4 x 10^-5), indicating that the normative twin captures cross-site-transferable structure rather than a scanner- or pipeline-specific artefact. Our findings suggest that cardiac phenotyping can move beyond predefined measurements toward data-driven dynamic phenotypes that better capture the heterogeneity underlying cardiovascular disease risk, even among individuals whose ejection fraction appears normal.

Counterfactual medical images are useful only when a requested clinical change is introduced without silently altering patient characteristics that should remain fixed. Existing diffusion editors can produce convincing images, but realism or target-classifier flips do not establish that an edit is localized, diagnosis-specific, anatomically faithful, or useful on real clinical data. We introduce PATHEDIT, a state-factorized counterfactual editing framework that decomposes an abdominal CT representation into pathology, anatomy, acquisition, and residual context before performing a clinically specified latent intervention. A pathology editor modifies only the target factor, while non-target factors are explicitly copied and preservation losses penalize collateral drift. The edited state is decoded jointly into an image, a report level description, and diagnostic readouts. Weak supervision from lesion masks, organ segmentations, and naturally occurring longitudinal lesion changes anchors edit direction and spatial support without requiring pixel-aligned counterfactual ground truth. We evaluate four properties that are often conflated in medical image editing: target edit success, non-target preservation, anatomical identity, and real-data utility. In our evaluation protocol, PATHEDIT achieves 94.2% target success while reducing mean off-target diagnostic change to 3.0%, yields substantially stronger alignment with observed lesion changes, and improves rare-lesion recognition and shortcut robustness on held-out real CT. These results motivate astricter view of counterfactual medical imaging: a useful edit must prove not only what changed, but also what did not.

Accurate prediction of tumor recurrence in brain tumor patients following surgery is essential for optimizing adjuvant therapy, response assessment, and surveillance regimen. While MRI remains the gold standard for surveillance, integrating patient-specific clinical context may inform recurrence prediction. Traditional multimodal deep learning approaches often incorporate clinical data via simple fusion, failing to fully capture the semantic interdependencies between visual features and clinical context. Trained on over 5,000 scans from approximately 400 pediatric low-grade glioma subjects and validated across three institutional cohorts, including one clinical trial cohort, our experiments demonstrate incremental performance gains when progressing from vision-only to clinical-vision to a vision-language approach. Our results indicate that converting structured clinical covariates into natural language text allows for more effective synthesis of multimodal data, while providing a platform for incremental addition of clinical context without extending model complexity. We demonstrate that our proposed VLM architecture offers a promising direction for neuro-oncological prognosis by effectively encoding imaging cues and clinical context, with potential applicability to other longitudinal prognosis tasks.

Maternal healthcare prediction systems often suffer from algorithmic biases due to socio-economic disparities and imbalanced datasets, limiting their effectiveness for equitable healthcare policymaking. This paper introduces MaternaAI, a fairness-aware and explainable learning framework designed to enhance maternal healthcare predictions in Kerala, India. The framework focuses on three critical health indicators:(1) Tetanus Toxoid (TT) booster uptake,(2) immunization coverage rates, and (3) the percentage of pregnant women completing four or more Antenatal Care (ANC) visits. To address fairness, we propose Adaptive Equity Score Optimization (AESO), a novel optimization algorithm that dynamically integrates fairness constraints into model training. AESO is model-agnostic and adapts group equity weights in response to real-time disparities. We integrate SHAP, LIME, and feature permutation techniques for explainability, enabling transparent global and local interpretation. Empirical results demonstrate that MaternaAI significantly improves fairness metrics and model accuracy across diverse machine learning and deep learning models, offering interpretable and equitable decision support for public health stakeholders.

Clinical brain MRIs from pediatric health systems represent a viable resource for modeling early neurodevelopmental trajectories and studying neurodevelopmental risk in real-world populations. However, a limitation to date has been the performance of existing segmentation tools for measuring various brain phenotypes in clinical scans. In particular, many tools underperform in infant scans due to morphological and physical changes such as rapid myelination. Here, we introduce ClinSeg: a robust segmentation approach tailored to early-life clinical MRIs with variable orientation, resolution, and contrast. We leverage existing registration and synthetic data generation tools to construct a training corpus for a 3d U-Net spanning anatomical and contrast diversity, including scans with morphological abnormalities from a pediatric hospital. Validated against manual segmentations, ClinSeg outperforms existing models in infancy while matching them in childhood and adolescence. Finally, ClinSeg enables the construction of reference brain growth trajectories in 11,699 individuals from 0-21 years of age, leading to the detection of more nuanced age-related findings in clinical groups.