Skip to content

🏥 Medical Imaging

🧠 NeurIPS2026 · 77 paper notes

📌 Same area in other venues: 🎞️ ECCV2026 (112) · 📷 CVPR2026 (182) · 🔬 ICLR2026 (87) · 🧪 ICML2026 (28) · 🤖 AAAI2026 (78) · 🧠 NeurIPS2025 (78)

🔥 Top topics: Medical Imaging ×39 · Multimodal/VLM ×8 · Diffusion Models ×5 · Alignment/RLHF ×4 · Face & Gaze ×3

A Steerable Deep Network for Model-Free Diffusion MRI Registration

Presents the first model-free nonrigid diffusion MRI registration framework based on an SE(3)-equivariant steerable UNet, framing registration as an equivariant diffeomorphism over position-and-orientation space with a Fourier-space MMD loss that aligns ensemble average propagators directly from raw data without derived models or explicit reorientation.

Aegis: Generative Gradient Masking for Privacy-Preserving Medical Federated Learning

Aegis adds a task-relevant synthetic-image gradient update after each client-local epoch to increase sample mixing in linear leakage; it reduces the measured reconstruction rate to 9.38%–11.63% across three MedMNIST modalities, without providing zero-leakage or differential privacy guarantees.

Aligning LLMs with Biomedical Knowledge using Balanced Fine-Tuning

Addressing the unique uncertainty structure in biomedical text where dense runs of low-confidence tokens encode crucial entities and mechanistic chains rather than stylistic noise, Balanced Fine-Tuning (BFT) employs local group-normalized token reweighting and sequence-level hard-sample reallocation, delivering superior reasoning and a resilient initialization for sparse-reward RL.

Ask, Answer, and Detect: Role-Playing LLMs for Personality Detection with Question-Conditioned Mixture-of-Experts

ROME introduces an Ask–Answer–Detect framework that leverages LLM role-playing to generate questionnaire-grounded intermediate answers from user posts offline, followed by a question-conditioned Mixture-of-Experts (MoE) network and reliability-importance adaptive modulation to achieve state-of-the-art personality detection with lightweight local inference.

Assessing the robustness of heterogeneous treatment effects in survival analysis under informative censoring

Addressing informative patient dropout in clinical survival analysis, this paper introduces an assumption-lean partial identification framework with sensitivity functions alongside SurvB-learner, a doubly robust and quasi-oracle efficient meta-learner that reliably identifies treatment-responsive subgroups under censoring bias.

Beyond Sparse Captions: Aligning Slide-Level Text and Patch-Level Vision in Pathology

LIMIT (Locked Image Multiple Instance PreTraining) adapts a text encoder to a frozen patch-level visual backbone using report-guided cross-attention pooling and difficulty-weighted InfoNCE, yielding CONCH-Z trained on 335k routine diagnostic reports that sets state-of-the-art zero-shot performance across 22 pathology tasks.

Beyond ViT Tokens: Masked-Diffusion Pretrained Convolutional Pathology Foundation Model

Addressing the issue where Vision Transformer (ViT) patch tokenization disrupts spatial continuity and local cellular morphology, this paper proposes ConvNeXt Masked-Diffusion (CMD), a self-supervised generative pretraining framework using a purely convolutional ConvNeXt-UNet in pixel space conditioned on frozen foundation model semantics, achieving state-of-the-art cell-level dense prediction with minimal task-specific annotation.

B[FM]\(^2\): Brain Foundation Model via Flow Matching with SplitUNet

This paper introduces B[FM]\(^2\), an EEG foundation model that discards conventional discrete patching and masked reconstruction to pretrain directly on raw continuous multi-channel EEG waveforms via optimal-transport conditional flow matching (OT-CFM) and a time-electrode factorized SplitUNet, achieving state-of-the-art performance on 7 of 9 downstream benchmarks with only \(\approx 307\text{ h}\) of pretraining data while generating synthetic EEGs that board-certified neurologists cannot distinguish from real clinical recordings (\(\kappa = -0.096\)).

BigCell: Generating Gigapixel Whole-Slide Images

BigCell introduces the first unified framework capable of natively generating gigapixel whole-slide histopathology images (WSIs), combining a progressively trained 2048-resolution flow-matching Transformer, pathology foundation model embedding-guided sliding-window inference, adaptive CFG velocity caching, and a text-conditioned blueprint generator (CGM).

BrainPro: Process-conditioned Pretraining with Topographic Retrieval for Large EEG Models

Addressing montage heterogeneity and the neglect of experimental protocol context in EEG foundation models, BrainPro couples a universal topographic spatial filter retrieval mechanism with process-conditioned multi-branch masked pretraining, outperforming existing baselines across nine public BCI benchmarks.

BrainWorld: A Structural-Prior-Conditioned Generative Model for Whole-Brain 4D fMRI Dynamics

BrainWorld formalizes whole-brain 4D fMRI generation as a structural-prior-conditioned spatiotemporal latent diffusion process, leveraging subject-level sMRI to guide future fMRI denoising and achieving stable 400-frame autoregressive rollouts alongside superior downstream transfer representations.

Bridging the Modality Bottleneck in Pathology MIL Through Virtual Molecular Staining

To overcome the fundamental modality bottleneck where frozen pathology foundation models fail to capture molecular states from morphology alone, MIST introduces a plug-and-play projection layer that derives cross-modal molecular prototypes from training-only paired spatial transcriptomics, delivering consistent gains across 23 downstream tasks and 8 MIL aggregators using pure H&E at inference.

C3VD-DEFCOL: A Deformable Colonoscopy Dataset with Time-Resolved 3D Ground Truth and Realistic Appearance

Addressing the trade-off between unavailable 3D ground truth in vivo and unrealistic synthetic appearances, C3VD-DEFCOL introduces a benchmark featuring parameterized peristaltic deformation paired with a depth-conditioned video diffusion model (LTX-2.3 Sim2Real), providing 110 video sequences (9,240 frames) with time-resolved 3D ground truth and in vivo mucosal appearance.

CasePlay: Self-Play Reinforcement Learning from Case Reports for Medical Reasoning

To overcome knowledge-point fixation and difficulty-quality mismatch in document-grounded self-play RL over medical corpora, CasePlay introduces a knowledge-conditioned proposer for comprehensive case evidence coverage and a rubric-judge reasoner providing quality-aware feedback across clue integration, reasoning depth, and option quality, yielding substantial gains across seven medical reasoning benchmarks.

CASL-VAE: Learning Structured Latent Variables from Unpaired Data for Semi-supervised Clustering and Paired Sample Generation

Addressing the challenge of unobserved cross-domain pairings and target-specific heterogeneity in biomedical comparative studies, CASL-VAE factorizes continuous common and hierarchical salient latent variables under posterior decoupling, enabling unsupervised subtype discovery and faithful paired sample generation.

CD-RCM: Generalizable Continuous-Depth Novel View Synthesis for Reflectance Confocal Microscopy

Addressing the severe axial anisotropy in in vivo skin reflectance confocal microscopy (RCM) due to optical sectioning limits, CD-RCM introduces a feedforward, depth-queryable novel-view synthesis framework combining virtual camera axial geometry with a domain-adapted perceptual loss, achieving sub-second continuous-depth slice generation and isotropic 3D tissue visualization without per-stack optimization.

Cephalonauts-One: A Deep fMRI Dataset for Decoding Naturalistic Speech in the Human Brain

This paper introduces Cephalonauts-One, a deep dataset featuring 30 hours of whole-brain 3T fMRI recordings per participant across three subjects listening to naturalistic podcasts, alongside an end-to-end audio segment retrieval benchmark and scaling law analysis.

CRAFT: Causal Responsibility and Failure Tracing in Medical Vision Language Models

Proposes CRAFT, a structural causal framework that utilizes counterfactual clean-state interchange to deconfound observational attributions and isolate spatially disjoint attention heads mediating arbitration and brake failures in medical VLMs, achieving training-free failure mitigation without sacrificing diagnostic fidelity.

DDx-TRACE: A Benchmark for Medical Diagnostic Trajectories in VLMs

DDx-TRACE establishes a physician-adjudicated benchmark of 211 neuroradiology cases that reformulates multimodal medical diagnosis as an interactive, sequential evidence-acquisition trajectory under partial information, revealing that frontier VLMs' high endpoint diagnosis scores frequently conceal essential-evidence omissions, erratic exam ordering, and poorly calibrated belief updates.

DeepArrhythmia: Segment-Contextualized ECG Arrhythmia Classification via Selective Evidence Acquisition

DeepArrhythmia produces R-peak-aligned beat labels within 10-second ECG segments and uses initial segment confidence to decide whether to acquire numerical and morphology evidence, demonstrating the value of context and physiological grounding across four datasets without outperforming always-rich inference on every metric.

EEG Benchmarking Needs a Task Specification Layer: NeuroDoc for Rulebook-Guided, Executable Benchmark Construction

Addressing the core obstacle that heterogeneous EEG datasets lack unified task definitions and cannot be freely redistributed due to privacy and licensing constraints, this paper introduces a rulebook-guided task specification language pairing human-readable documents with executable kernels, supported by the NeuroDoc and NeuroAudit toolchain, releasing a community-audited corpus of 53 completed datasets and 245 tasks.

Extremely Sparse-View Computed Tomography from 2D Projections via Pose-Aware Diffusion Priors

Proposes a train-sparse / infer-sparse paradigm for extremely sparse-view CT (xSV-CT) by learning a multi-scale pose-aware diffusion prior from only a single 2D projection per sample, guided at inference by shared-volume reprojection consistency and discrete Radon range condition surrogates to achieve state-of-the-art novel view synthesis and high-fidelity 3D reconstruction without dense volume supervision.

Fine-Tuning Language Encoding Models on Slow fMRI Improves Prediction for Fast ECoG

Fine-tuning the self-supervised speech model WavLM on slow non-invasive fMRI data significantly enhances encoding prediction performance for fast invasive ECoG brain responses across unseen subjects and stimuli, scaling logarithmically with fMRI training data.

FlatClip: Reusing Image Foundation Models for fMRI Representation Learning via Cortical Flatmaps

FlatClip projects and rasterizes 4D fMRI cortical activity into anatomy-linked 2D flatmaps and extracts representations using a completely frozen out-of-domain SigLIP2 image encoder, demonstrating that training only a lightweight probe substantially outperforms specialized ROI foundation models and highlighting cortical spatial arrangement and task-region selection as decisive design factors.

Flow Matching with In-Context Priors for Out-of-Distribution Brain Dynamics

Presents a per-timestep conditioned diffusion transformer (DiT) trained with a flow matching objective, combining compositional language descriptions with self-supervised spatial priors to achieve zero-shot out-of-distribution generation and counterfactual simulation of whole-cortex fMRI dynamics.

Focus, Align, and Diffuse: Time-Series-Aware Keyframe Diffusion for Cardiac Dynamic Synthesis from Sparse Observations

CardioFAD reformulates one-shot sparse cardiac dynamic synthesis as a recoverability-alignability anchor discovery problem, self-discovering motion-salient keyframes from ECG time-series guided by the Trajectory Sufficiency Principle and Anchor Alignment Principle, and performing high-fidelity, functionally consistent full-cycle generation via a unified two-stage diffusion model.

FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs

FOCUS connects harmonized disease labels from ten fundus datasets and the prediction interfaces of three foundation-model families to a shared evaluation layer, finding that ranking, calibration, and subgroup behavior do not align, while MLLM supervised fine-tuning primarily improves binary decisions and calibration with little average AUROC gain.

Gate-Based Feature Importance in Simulation-Based Inference for EEG: A Cautionary Tale

In a Jansen–Rit EEG neural posterior estimation setup, this paper systematically compares jointly trained differentiable feature gating against a gold-standard retrain-based ablation benchmark, revealing a significant negative correlation (\(\rho \approx -0.61\)) and demonstrating that predictive gating fails to faithfully capture causal feature importance despite negligible accuracy loss.

GLINT: Sparsely Gated Vision-Language Alignment for Fine-Grained Radiology Representations

GLINT resolves the scale mismatch between local clinical findings and global report supervision via Sparsely Gated Alignment (SGA) with a decoupled Sigmoid gate space and Dense Feature Regularization (DFR) anchored to frozen SSL teachers, delivering state-of-the-art fine-grained representations across 2D CXR and 3D CT, and demonstrating the first zero-shot free-text 3D CT volume segmentation without mask supervision.

GLINT: Sparsely Gated Vision-Language Alignment for Fine-Grained Radiology Representation Learning

To tackle the severe scale mismatch between focal lesions and global reports in radiology VLMs, GLINT introduces Sparsely Gated Alignment (SGA) with unnormalized Sigmoid gates and Dense Feature Regularization (DFR) against frozen SSL teachers, achieving state-of-the-art zero-shot grounding and the first mask-free zero-shot 3D CT segmentation.

GRAPE: Graph-Augmented Prototype Explanations for Interactive Medical Image Diagnosis

Addressing prototype-based medical image classifiers' limitations of treating findings as independent, silently amplifying unsafe physician feedback, and requiring full retraining for new findings, GRAPE unifies a data-driven graph attention head, prototype-variance spatial safety checks, and frozen BioViL-T open-vocabulary anchoring, delivering a +13.8 pp macro-F1 gain and intercepting 85% of erroneous annotations at +1 ms latency.

GRASP: Guided Residual Adapters with Sample-wise Partitioning

GRASP mitigates catastrophic mode collapse of tail classes in text-to-image flow matching transformers by deterministically partitioning the conditioning space (using frequency/healthy priors or text embeddings) and routing samples to dedicated feedforward residual adapters, preventing head-versus-tail gradient interference while leaving the flow-matching objective and inference sampler completely non-invasive.

Handwriting Decoding as a Challenging Motor Task for EEG Foundation Models

By identifying critical oculomotor, visual, and block-design confounds in existing EEG handwriting datasets, this work introduces the confound-free, cross-session LOVW-8.5K benchmark, revealing that state-of-the-art EEG foundation models are outperformed by compact specialist decoders and demonstrating that movement onset uncertainty and test-time signal-to-noise ratio (SNR) form the primary bottlenecks.

HyperNSDE: Personalized Neural SDEs for Joint Static–Longitudinal Clinical Data Generation

HyperNSDE introduces a continuous-time generative framework that conditions a latent Neural SDE on patient static covariates via a hypernetwork, combined with a deterministic–stochastic path decomposition, non-adversarial signature-kernel scoring, and latent-state-dependent point-process intensities to jointly synthesize heterogeneous static variables, irregular longitudinal trajectories, and informative observation times.

Information Bottleneck-Guided Adaptive Hypergraph Transformer for Brain Disease Diagnosis

IBAHGT combines adaptive message passing on a fixed nearest-neighbor hypergraph with a parallel global Transformer, then uses node-level fusion and information-bottleneck auxiliary training to improve fMRI brain-network classification, achieving 77.6% / 86.3% accuracy on randomly split ABIDE / ADNI, although its mutual-information derivations contain important conditional and sign issues.

Knowledge Transfer Scaling Laws for 3D Medical Imaging

Addressing the challenge of unprincipled heuristic data mixing in multi-modal 3D medical foundation models, this paper introduces a transfer-aware scaling law that models domain-specific scaling exponents alongside an asymmetric cross-domain transfer matrix, enabling closed-form compute-optimal data allocation that outperforms standard heuristics on MAE reconstruction and diverse downstream clinical tasks.

L-FAME: Longitudinal Focused Attention Meditation EEG Dataset and Benchmark

L-FAME is the first 6-week longitudinal, 64-channel EEG dataset tracking 74 participants across three distinct focused attention meditation practices (breath focus and two mantra techniques), establishing a standardized benchmark across cognitive state decoding, fine-grained technique classification, and few-shot cross-session adaptation to systematically dissect temporal neural drift.

LEAF: Language-EEG Aligned Foundation Model for Brain-Computer Interfaces

LEAF proposes a foundation model for EEG–language alignment that unifies self-supervised joint spectral-temporal reconstruction (STR) with an instruction-conditioned Q-Former (IQF), achieving state-of-the-art decoding performance across 16 downstream BCI benchmarks spanning five distinct task paradigms.

Learning What Not to Impute: An Uncertainty-Aware Diffusion Framework for Meaningful Missingness

Addressing the challenge of mixed "observation-induced missingness" and "semantically valid intrinsic missingness" in tabular data, this paper proposes Diff-Joint, a joint diffusion framework that couples conditional generation with posterior uncertainty clustering to selectively impute missing values while preserving meaningful missingness.

Learning When to Collaborate: Selective Multi-Agent Medical Reasoning via Uncertainty-Aware Routing

SMART-Med addresses the prohibitive cost and privacy risks of cloud-dependent, dense multi-agent medical systems by introducing a fully local framework driven by uncertainty-aware routing, fine-tuned via GRPO on a hierarchical reasoning dataset MedVR-3K with certainty-aware accuracy and multimodal quality rewards.

MedHorizon: Towards Long-context Medical Video Understanding in the Wild

MedVTok: A General-Purpose Medical Visual Tokenizer

MedVTok is a general-purpose medical visual tokenizer that preserves through-plane volumetric continuity via slice differential regularization and infuses clinical semantics via bidirectional multi-expert alignment, unifying 2D images and 3D volumes across 9 imaging modalities into a single token interface.

MedVTok: A General-Purpose Medical Visual Tokenizer

MedVTok is a general-purpose medical visual tokenizer that maps 2D images and 3D volumetric scans across diverse modalities into a unified token space by coupling slice differential regularization for through-plane anatomical continuity with multi-expert representation alignment for cross-modal clinical semantics, setting new state-of-the-art results across 24 downstream benchmarks.

MentalHospital: A Virtual Environment for Evaluating Psychiatric Clinical Encounters

This paper introduces MentalHospital, the first EHR-grounded virtual evaluation environment instantiating the full S.O.A.P. psychiatric encounter workflow, alongside MentalEval, a specialist-aligned suite of evaluators, revealing substantial gaps between current LLMs and human clinicians in psychiatric competence and mental status assessment.

MID: Mask-Image Distributional Divergence for Evaluating Medical Image Segmentation

Introduces MID (Mask-Image Distributional Divergence), the first distribution-level unsupervised evaluation metric for medical image segmentation, which embeds paired images and class-wise binary masks into a joint Gaussian latent space using a pretrained medical vision encoder to compute the squared 2-Wasserstein distance without requiring paired ground-truth annotations.

MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding

MindAlign establishes a tri-modal contrastive alignment framework bridging EEG, visual images, and LLM-generated descriptive text, combining spatio-temporal masked reconstruction pre-training with a specialized encoder to achieve state-of-the-art 200-way zero-shot visual decoding.

MIRAGE: Adaptive Multimodal Gating for Whole-Brain fMRI Encoding

Presents MIRAGE, a whole-brain fMRI encoding framework that pairs frozen Qwen3-Omni multimodal foundation models with learnable cross-attention layer gating poolers and a shared temporal Transformer, setting a new state-of-the-art in naturalistic brain response prediction and out-of-distribution generalization on Algonauts 2025.

Modeling Whole-Slide Images as Dynamic Tumor Microenvironment Fields

TMEvolve models pathology slides as repeatedly updated latent region fields, learning slide representations through intra-region diffusion and concept-guided directed boundary interactions; it achieves the highest mean C-index on three of four survival cohorts, but its “evolution” denotes representation refinement rather than actual tumor chronology.

MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI

Presents MosaicMRI, the largest open-source collection of raw musculoskeletal (MSK) MRI measurements to date (3,150 multi-coil volumes and 89,830 slices spanning 10 anatomy classes and 0.55T/1.5T field strengths), establishing accelerated reconstruction benchmarks and demonstrating that mixed-anatomy training consistently outperforms anatomy-specific baselines under realistic domain shifts.

Multimodal LLMs Outperform Pathology Foundation Models in Cross-Domain Histological Similarity

Using three-image similarity choices across six histopathology cohorts, MOSAIC directly contrasts same-class cross-domain candidates with different-class same-domain candidates and finds that zero-shot multimodal large language models usually resist acquisition-context shortcuts better than Euclidean distances between pathology foundation model embeddings, without establishing clinical diagnostic competence or reliable cross-domain performance.

NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces

NeuroAtlas establishes a comprehensive benchmark of 42 datasets and ~260k hours of EEG across clinical neurology and BCIs, evaluating 20 foundation and supervised models under a unified frozen-probe protocol to expose substantial intra-domain rank shifts, marginal gains over general time-series models, and susceptibility to non-neural confounds.

NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces

NeuroAtlas establishes a massive benchmark spanning 42 datasets, ~260k hours of EEG, and 20 models across clinical domains and BCIs under a unified frozen-probe protocol, revealing that EEG foundation models do not reliably outperform generic time-series models and that conventional aggregate metrics systematically mask clinical failure modes.

Nonparametric Distribution Matching for Self-Supervised Whole-Slide Image Condensation

This paper reformulates whole-slide image (WSI) condensation as an optimal transport distribution-matching problem under a fixed frozen encoder representation lens and proposes NICER, a nonparametric framework that decouples local pattern exploration from global concept condensation to achieve slide-adaptive prototype capacity and major downstream performance gains.

Paleoinspired Vision: From Exploring Colour Vision Evolution to Inspiring Camera Design

Presents a paleoinspired computational vision framework that abstracts retinal visual transduction into a Gaussian-parameterized 1×1 convolutional opsin layer with micro-step mutation regularisation, reconstructing millions of years of vertebrate colour vision evolution in seconds on GPU and guiding manufacturing-friendly spectral response design for Martian exploration and cancer detection cameras.

Pan-FM: A Pan-Organ Foundation Model with Saliency-Guided Masking for Missing Robustness

Pan-FM tackles the dominant-organ shortcut learning bias and non-random organ missingness in multi-organ biomedical self-supervised learning via a unified ViT backbone with cross-attention patch encoding and attention-based Saliency-Guided Masking (SGM), achieving robust whole-body representations across 7 organ systems.

Positive-Unlabeled Preference Optimization for Chest X-Ray Report Generation

Addressing the issue of omission noise where clinically present findings in retrospective chest X-ray reports are left unmentioned, this paper introduces PU-DPO, a positive-unlabeled direct preference optimization framework that combines LLM-based two-stage contrastive edits with an unbiased risk estimator to recover hidden positives and improve clinical detection rates.

QUTCC : Quantile Uncertainty Training and Conformal Calibration for Imaging Inverse Problems

QUTCC embeds continuous quantiles into an Attention U-Net for simultaneous quantile regression and performs independent risk-controlled conformal calibration in the input quantile space, delivering non-linear spatially-adaptive scaling with valid marginal coverage guarantees while pinpointing model hallucinations in accelerated MRI, quantitative phase imaging, and denoising.

RAM-H1200: A Unified Evaluation and Dataset on Hand Radiographs for Rheumatoid Arthritis

Presents the first multi-center benchmark and dataset combining whole-hand bone instance segmentation, pixel-level bone erosion delineation, and clinically grounded SvdH joint scoring, revealing that micro-lesion quantification remains a critical bottleneck compared to mature anatomical modeling.

Reporting Practice Matters: The Impact of Reference Choice on Chest X-ray Report Evaluation

Reveals that variations in radiologist reporting practices significantly confound automated chest X-ray report generation evaluation, introduces an expert-validated four-dimensional taxonomy alongside a clinical-preserving rewriting method (ReRef), and demonstrates that mainstream metrics fail to decouple clinical accuracy from reporting style, to the extent of reversing leaderboard rankings.

Scaling Multi-Teacher Distillation for Digital Pathology

Addressing the scaling failure of conventional multi-teacher distillation when broadening the teacher pool, this paper proposes teacher-specific tokens, structured attention masking, tail MoE layers, and InfoNCE contrastive regularization to distill 10 SOTA pathology foundation models into a lightweight OMNI encoder family (7M-330M), achieving up to 10× compute efficiency while outperforming all teachers.

SheafStain: Sheaf-Theoretic Schrödinger Bridge for Spatially and Biologically Coherent Virtual Staining

SheafStain adds neighborhood VFM spatial conditioning, overlap-consistency losses, and pathology supervision to an unpaired Schrödinger bridge, then uses adaptive extra patches and weighted stitching to generate IHC from H&E, substantially reducing seams on 1024×1024 assembled regions from two breast pathology datasets without guaranteeing diagnostic correctness.

SliceWorld: A Predictive and Controllable World-State Model for CT Report Generation

By framing 3D axial CT slice sequences as causally evolving world states decomposed into anatomy, lesion, and uncertainty factors, SliceWorld leverages multi-step future-slice feature prediction and counterfactual lesion intervention to significantly enhance CT report consistency and controllable analysis.

SPERA: Spherical Prior EEG Foundation Model with Geometry- and Frequency-Aware Latent Prediction

SPERA is an electroencephalography (EEG) foundation model based on the joint-embedding predictive architecture (JEPA) that incorporates a biophysical Legendre-polynomial spherical prior and relational spectral regularization, achieving state-of-the-art transferability across diverse clinical, cognitive, and brain-computer interface (BCI) tasks.

Standing on the Shoulders of Giants: Rethinking EEG Foundation Model Pretraining via Multi-Teacher Distillation

Addressing the data scarcity and low signal-to-noise ratio in electroencephalography (EEG), this paper replaces conventional self-supervised masked reconstruction with Multi-Teacher Distillation Pretraining (MTDP), fusing visual and time-series foundation models via a two-stage masked latent denoising objective to train a compact EEG student that surpasses full-data self-supervised SOTA with only 25% of the pretraining data.

STREAM: Stochastic Riemannian Flow Matching with Anisotropic Decoder for Digital Histopathology Image Generation

Addressing the conditioning-dominated diversity and semantically poor VAE latents in state-of-the-art histopathology synthesis, STREAM directly adopts the vision foundation model (VFM) feature space as the generation hypersphere \(\mathcal{S}^{d-1}\), combining tangent-Gaussian Brownian-bridge Riemannian flow matching with an SVD-informed anisotropic decoder to achieve state-of-the-art unconditional generation.

SurgVista: Long-Horizon Surgical World Modeling with Plausible Instrument-Tissue Dynamics

SurgVista addresses spatial interaction incoherence and autoregressive temporal fidelity collapse in surgical video generation through latent trajectory-based Deformation Consistency Regularization and online residual-augmented Drift Adaptation Training, enabling physically plausible, stable surgical rollouts up to 800 frames.

Symb-xMIL: Symbolic Explanations for Multiple Instance Learning in Digital Pathology

Addressing the inability of spatial heatmaps in multiple instance learning (MIL) to explain higher-order interactions across tissue regions, Symb-xMIL introduces a post-hoc symbolic explanation framework that maps MIL predictions to human-readable Boolean logic rules, successfully identifying ground-truth multi-instance logic, exposing hidden Clever Hans errors in clinical tumor detection, and refining prognostic stratification in cancer cohorts.

Text-Guided Medical Image Segmentation via Residual Aggregation and Dense Alignment Representations

RADAR presents a parameter-efficient framework for text-guided medical image segmentation that explicitly decouples BiomedCLIP semantic anchoring from MedSAM2 promptable mask decoding, using text-conditioned residual depth aggregation (GRDA) and structured dense alignment (SDRA) to achieve state-of-the-art segmentation with only 0.74M trainable parameters.

Tokenizer-Generator Coupling in Medical Image Generation

This controlled factorial study treats the tokenizer–generator–sampler triple as its experimental unit: across 70 generation cells on ChestMNIST-64, reconstruction quality does not reliably predict generation quality, quantizer rankings depend on the generator, and validation-selected reduced-step sampling lowers LFQ-1024 + D3PM FID-192 from 0.44 to 0.09.

U-Bench: A Comprehensive Understanding of U-Net through 100-Variant Benchmarking

Benchmarking 100 U-Net variants across 28 datasets and 10 medical imaging modalities reveals that over 80% of reported in-domain gains lack statistical significance, while introducing U-Score as a deployment-oriented metric and a feature-driven advisor agent to guide real-world architecture selection.

UniPro: Unified Multi-Mode Medical Image Segmentation from 2D Images to 3D Volumes via Propagation

UniPro unifies semantic, in-context, interactive, and volumetric propagation segmentation into a single slice-based 2D encoder–decoder architecture by treating slice propagation as intra-volume reference-conditioned in-context prediction, achieving coherent 3D segmentation from sparse 2D interactions with zero extra 3D parameters.

ViCoR: Estimating Visual Necessity via Counterfactual Residuals for Multimodal Medical Data Selection

Addressing modality imbalance and language shortcuts in Med-VLM fine-tuning, ViCoR isolates image-dependent optimization residuals using neutral visual controls, combining clinical supervision calibration with diversity-aware selection to construct a 20% subset that outperforms full-data fine-tuning across generalization and multi-disease benchmarks.

VIPER: An Expert-Curated Benchmark for Vision-Language Models in Veterinary Pathology

VIPER is the first expert-curated benchmark for multimodal vision-language models in rodent toxicologic pathology, establishing a rigorous 1,251-question testbed across 7 organ systems and demonstrating a substantial domain gap for human-specialized models and an over-diagnosis vulnerability in frontier models.

WASP: Weakly Aligned Spatiotemporal Pairs for Fetal Brain MRI–Ultrasound Learning

Addressing the clinical unavailability of same-subject same-day paired fetal brain MRI–ultrasound (US) scans, WASP leverages routine clinical metadata of Gestational Age (time) and diagnostic planes (space) via entropic Optimal Transport and barycentric projection to adapt frozen ultrasound backbones to MRI without fine-tuning.

Where to Look Matters: Rethinking Sub-Volume Sampling in 3D Medical Self-supervised Learning

Addressing the issue where uniform random cropping ignores spatial redundancy and anatomical heterogeneity in 3D medical SSL, VolumeProbe provides a training-free plug-and-play framework that constructs an offline probe bank based on anatomy-awareness, informativeness, and diversity, substantially boosting downstream segmentation, classification, and registration across VoCo, S2DC, and MedGMAE.

Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It

Addressing the steep performance drops of 3D biomedical image segmentation under domain shifts and the theoretical performance ceilings of pure invariant representations, MaskGen concatenates source image intensities with frozen foundation model representations and regularizes their combination via stochastic channel view masking, achieving state-of-the-art generalization without any test-time adaptation.

XDecomposer: Learning Prior-Free Set Decomposition for Multiphase X-ray Diffraction

To tackle severe peak overlap and unknown phase mixtures in multiphase powder X-ray diffraction (PXRD), XDecomposer formulates multiphase analysis as a blind source separation set-prediction task, combining self-supervised global pretraining, phase-query latent modulation, and physics-consistent continuous masking to cleanly disentangle single-phase signals and abundance ratios in one shot without candidate phase priors.