Root canal therapy is a prevalent dental procedure for treating pulp infections; however, its complexity due to the intricate variability of root canal anatomies frequently results in procedural errors, excessive tooth structure loss, and suboptimal outcomes, especially in multi-rooted teeth. Although Guided Endodontics enhances precision and predictability, it depends on manual planning using non-dedicated software, rendering the process labor-intensive, subjective, and confined to single-rooted teeth, thereby limiting broader adoption. In this study, we introduce EndoPlanner, an adaptive and pioneering framework that streamlines preoperative planning for Guided Endodontics solely from cone-beam computed tomography scans, seamlessly integrating automated dental segmentation, endodontic landmark detection, access cavity design, and surgical template generation. Specifically, we propose a unified bottom-up multi-peak encoding and graph-based decoding paradigm for root canal landmark localization and topology recognition, augmented by an anatomy-aware loss for targeted heatmap optimization and a self-supervised inference-time refinement strategy to bolster generalization across diverse root canal morphologies. Additionally, a tunable, geometry-guided algorithm is devised for automatic drill path planning, incorporating minimally invasive principles. Extensive evaluations affirm the superior performance and robustness of our method, yielding an average MRE of 0.767 mm, with SDRs of 74.9% within the strict 1.0 mm safety margin and 94.3% within the broader 2.0 mm tolerance for landmark localization, while attaining clinically acceptable standards for access planning. Deployed as a 3D Slicer extension, it completes preoperative preparation in an average of ∼4 min, reducing time by up to 96.05% relative to traditional workflows. Moreover, successful in-vitro experiments and clinical cases validate the efficacy of patient-specific access solutions facilitated by our efficient, reproducible, and minimally invasive planning schemes, highlighting their potential for routine practice, particularly among less-experienced practitioners.
Functional magnetic resonance imaging (fMRI) enables noninvasive investigation of brain function, while short clinical scan durations, arising from human and non-human factors, usually lead to reduced data quality and limited statistical power for neuroimaging research. In this paper, we propose BrainCast, a novel spatio-temporal forecasting framework specifically tailored for whole-brain fMRI time series forecasting, to extend informative fMRI time series without additional data acquisition. It formulates fMRI time series forecasting as a multivariate time series prediction task and jointly models temporal dynamics within regions of interest (ROIs) and spatial interactions across ROIs. Specifically, BrainCast integrates a Spatial Interaction Awareness module to characterize inter-ROI dependencies via embedding every ROI time series as a token, a Temporal Feature Refinement module to capture intrinsic neural dynamics within each ROI by enhancing both low- and high-energy temporal components of fMRI time series at the ROI level, and a Spatio-temporal Pattern Alignment module to combine spatial and temporal representations for producing informative whole-brain features. Experimental results on resting-state and task fMRI datasets from the Human Connectome Project demonstrate the superiority of BrainCast over state-of-the-art time series forecasting baselines. Moreover, fMRI time series extended by BrainCast improve downstream cognitive ability prediction, highlighting the clinical and neuroscientific impact brought by whole-brain fMRI time series forecasting in scenarios with restricted scan durations.
Precise survival prediction from Whole Slide Images (WSIs) is pivotal for precision oncology yet remains challenging due to the gigapixel resolution and label scarcity. Current approaches are often hindered by two fundamental limitations: the domain misalignment between natural-image-pretrained vision foundation models and histopathological data, and the lack of high-order correlations modeling among sparsely distributed prognostic regions. In this paper, we propose a dual-stage Prognostic Saliency-driven Hypergraph Neural Network (ProSH-Net) for WSI-based survival prediction. A parameter-efficient distillation-based domain-adaptive pre-training paradigm is proposed to fully leverage the representation capability of vision foundation models while ensuring sensitivity to fine-grained morphological patterns. Subsequently, a saliency-aware hypergraph neural network is proposed to orchestrate feature aggregation. Different from traditional graphs, our method constructs hyperedges guided by semantic prototypes and spatial priors, explicitly modeling multi-to-multi interactions among patches. By injecting patch-level saliency scores into the message-passing mechanism, ProSH-Net effectively enhances critical prognostic signals. Extensive experiments on public benchmarks demonstrate the superiority of our proposed method over state-of-the-art alternatives.
The cortical sulci of the posterior medial cortex play a crucial role in cognitive and behavioral functions. However, automatic labeling of these sulci is challenging due to severe class imbalance (i.e., only a small portion of the cortex is labeled) and high neuroanatomical variability. To tackle these challenges, we propose an effective spherical deformation-based sulcal labeling method that automatically enlarges the sulcal regions of interest (ROIs) while contracting unlabeled regions. This allows the proposed method to capture detailed sulcal patterns more effectively and suppress confounding analogous patterns. Specifically, we introduce a generic expansion function with a generalized condition for homeomorphic isotropic deformation. Based on this, we propose a parameterizable exponential expansion design which provides flexibility to fully accommodate non-uniform spatial densities of sulcal ROIs. We also propose an end-to-end SO(3)-equivariant model to enhance its expressive power while improving generalization capability under data scarcity. Furthermore, we offer an explicit regularization on the extent of deformation to provide the stable enlargement of ROIs while preventing the unintended contraction. In the experiments, we show that the proposed method outperforms baseline methods on both adult and pediatric cohorts, particularly in small and variable ROIs. The most notable improvement of 11.16% in Dice score was observed in the smallest sulcus, which is hominoid-specific and implicated in cognition and different neurodevelopmental and clinical disorders. Strong negative correlations between performance gains and ROI size confirm its suitability for labeling small sulci.
Visual prompt tuning (VPT) efficiently adapts foundation models for medical image classification but remains dependent on large labeled datasets. To overcome this, we explore semi-supervised learning (SSL) for label-efficient VPT. In VPT, while the backbone is well-regularized by large-scale pre-training to exhibit stable representations, the newly introduced prompts are initialized from scratch and remain sensitive to perturbations. Motivated by this discrepancy, we propose PromptMatch, a framework that explicitly regularizes the prompt space. By adversarially perturbing prompts on unlabeled data, it enforces prediction consistency across prompt variations to stabilize the adaptation process. Furthermore, we introduce (1) a Deep Prompt Supervision (DPS) module that applies deep supervision to prompt embeddings, utilizing previously neglected intermediate prompt information for SSL; and (2) an Instance-Specific Prompt (ISP) module, which generates instance-specific prompts to capture the unique characteristics of individual samples, overcoming the limitations of shared prompts in VPT. We evaluate PromptMatch on three real-world medical image classification datasets. Our extensive experiments demonstrate that it surpasses other SSL counterparts which apply perturbations to features or samples. PromptMatch reduces the reliance on labeled data, matching fully supervised performance with only 2.86% of the labels. Code is available at https://github.com/fdu-farm/PromptMatch.
Skin cancer is a prevalent and potentially fatal disease that emphasizes the need for accurate and interpretable diagnostic tools to improve patient outcomes. Although DL has advanced automated skin lesion analysis, most models rely solely on dermoscopic images and neglect the complementary clinical metadata. In this study, we propose FuseMD-XNet, a multimodal transformer-based framework that integrates dermoscopic images with structured patient metadata. The model employs intermediate fusion through feature concatenation and cosine similarity alignment, followed by adaptive certainty-guided fusion to dynamically weigh the modality contributions based on confidence estimates. To ensure transparency, the FuseMD-XNet incorporates multilevel explainability using ShapleyCAM, FinerCAM, and SHAP methods. The efficacy of FuseMD-XNet was validated on the PAD-UFES-20 dataset, where it achieved an overall mean diagnostic accuracy of 94.4±0.8% and a mean AUC of 95.9±0.5% across all lesion classes. The highest class-specific performance was observed for basal cell carcinoma (BCC), with an accuracy of 98.4±0.4% and an AUC of 98.7±0.3%, whereas melanoma achieved an accuracy of 97.9±1.8% and an AUC of 98.2±0.5%. On the ISIC 2019 dataset, FuseMD-XNet demonstrated strong generalization performance with an overall mean accuracy of 93.0±0.8% and a mean AUC of 94.6±0.6%, whereas melanoma achieved a class-specific accuracy of 94.7±1.6% and an AUC of 96.3±1.1%. Additionally, an integrated risk stratification module enabled personalized assessments validated by feature importance analysis. These results demonstrate the potential of FuseMD-XNet to improve the classification accuracy and interpretability of skin cancer.
Medical image segmentation foundation models (MedFMs) perform strongly across diverse imaging modalities, but their large size and computational demands hinder deployment in resource-limited clinical settings. Lightweight fine-tuning is impractical given the high training cost of MedFMs, and the efficiency benefits of integer inference remain underused. Post-training quantization (PTQ) is a promising alternative, yet existing PTQ methods fail on MedFMs because their heterogeneous weight distributions lead to severe accuracy degradation at low bit-widths. To address this challenge, we propose GPTQ-MedFM, a generalized post-training quantization framework tailored for medical foundation models. GPTQ-MedFM standardizes complex weight distributions under an ℓ∞-constrained normalization to produce quantization-friendly matrices, and then applies an efficient coordinate-descent solver to obtain high-fidelity low-bit representations. The method adds no extra computation or memory overhead at inference, enabling seamless deployment on medical edge devices. By explicitly modeling and mitigating quantization-induced errors, GPTQ-MedFM achieves state-of-the-art low-bit compression across six medical foundation models and nine imaging modalities - spanning nearly the full range of clinical imaging scenarios - and remains robust even with a single calibration sample. Its broad generalization and minimal calibration cost make GPTQ-MedFM a practical pathway for real-time AI-assisted diagnostics in resource-limited healthcare settings.
Deep learning models for colorectal polyp detection often face distribution shift issues due to dataset artifacts, variations in imaging equipment, and other clinical factors, which result in unreliable performance in real-world settings. To address this, several generalizable object detectors have been proposed, aiming to train models that can generalize across multiple target domains. However, existing approaches to generalizable detection typically rely on two main strategies: Unsupervised Domain Adaptation, which requires unlabeled data from the target domains, and Multi-Source Domain Generalization, which necessitates data from multiple source domains. These approaches are challenging to implement in clinical settings, where data privacy regulations limit the accessibility. This work proposes a novel approach for polyp detection in colonoscopy scenes under the more realistic setting of single domain generalization, called Generalizable Polyp Detection Transformer (GPDT). Our method leverages Vision Foundation Models for robust feature extraction and introduces a learnable-token-driven adapter mechanism to fine-tune these models with minimal additional parameters. This approach enables effective generalization across unseen clinical domains when only a single source domain is available for training. Extensive experiments on two multi-center polyp detection generalization benchmarks, PolypGen and REAL-Colon, show that GPDT achieves stronger performance compared with existing state-of-the-art methods across multiple target domains. Furthermore, we introduce an efficient variant, E-GPDT, that accelerates inference while preserving detection accuracy, yielding a favorable speed-accuracy trade-off under the evaluated edge-device configuration. E-GPDT is trained as a source-specific student distilled from adapter-enhanced GPDT teachers under the same single-source protocol, thereby connecting the high-capacity VFM-adapted detector and the lower-latency student detector within one unified pipeline. Our results demonstrate that adapting VFMs provides a promising pathway for improving cross-domain polyp detection generalizability.
Recent advancements in foundation models (FMs) have catalyzed a paradigm shift in medical image analysis. Unlike traditional task-specific artificial intelligence (AI) models, FMs leverage large-scale datasets to learn generalized representations that can be adapted to downstream clinical applications. Despite the rapid proliferation of FM research in medical imaging, there is a lack of unified synthesis that systematically maps the evolution of architectures, training paradigms, and clinical applications across modalities. To address this gap, this review provides a comprehensive and structured synthesis of FMs in medical image analysis by systematically organizing studies into two primary categories: vision-only foundation models (VFMs) and vision-language foundation models (VLFMs), based on their architectural foundations, training strategies, and downstream clinical tasks. A quantitative analysis was conducted on both VFMs and VLFMs to characterize temporal trends in dataset utilization and application domains, along with pooled performance and subgroup analyses. We also critically discuss persistent challenges, including cross-domain generalization, computational scalability, FM evaluation, fairness, and deployment. Finally, we identify key future research directions aimed at enhancing the robustness, interpretability, and clinical integration of FMs, thereby accelerating their translation into real-world medical practice.
Functional connectome (FC) fingerprinting is crucial for understanding individual cognitive patterns and advancing personalized medicine for neuro/psychiatric disorders by developing individual-specific biomarkers. However, existing FC fingerprinting methods oversimplify the complex and nonlinear nature of FC patterns. As a result, they fail to effectively extract individual-specific information from variability across different brain states, thereby limiting individual identification performance. To address this issue, we propose Hierarchical Attention and Hard Negatives-aware State Graph Contrastive Learning (HAHN-SGCL) model. HAHN-SGCL directly leverages brain states to generate intra- and inter-individual contrasts, effectively extracting individual-specific connectivity patterns for accurate identification across diverse states. Specifically, to fully extract individual-specific information across multiple topological levels of the FC, we designed a Hierarchical Graph Attention Network (HGAT) encoder. HGAT constructs a hierarchical graph with diverse topological perspectives and employs level-specific attention mechanisms to capture distinctive individual features. Additionally, to overcome the severe sample imbalance that hampers effective gradient propagation, we introduce a Hard Negatives-aware Strategy (HNS). HNS focuses on challenging negatives through Hard Negative Mining (HNM) and incorporating a corrective term, effectively avoiding early convergence plateaus. Extensive experiments demonstrate that our HAHN-SGCL model outperforms state-of-the-art methods. It also exhibits strong cross-task transferability, as evidenced by its robust performance in psychiatric disorder classification. The code of HAHN-SGCL is at https://anonymous.4open.science/r/HAHN-SGCL.
Automated detection of surgical errors can drastically increase safety in robotic-assisted surgery (RAS). Prior work has shown surgical error detection only as a binary task, which is insufficient and restrictive when multiple, co-occurring errors inherently arise in real RAS procedures. We thus present the first work to study multi-label surgical error detection (MLSED) and propose SEDCLIP, a novel vision-language adaptation paradigm that transfers pretrained knowledge from a contrastive language-image pretraining model to MLSED. The SEDCLIP architecture comprises: (1) global-local prompting, which aligns global image features and fine-grained patch features with learnable global and local text prompts, countering the single-label bias of contrastive pre-training; (2) a graph-attention temporal adaptor that models inter-frame dependencies, guided by a temporal-distance adjacency; and (3) random-basis low-rank adaptation, which aggregates fixed low-rank random bases modulated by learnable scaling, to bridge the natural-to-surgical domain gap with efficient fine-tuning. Additionally, to catalyse progress in MLSED, we release two first-of-their-kind, in-vivo datasets with frame-level error annotations: MLE-RAH for robotic-assisted hysterectomy (RAH) and MLE-RARP for robotic-assisted radical prostatectomy (RARP). Across both datasets, SEDCLIP yields overall improvements over competing methods on both multi-label and binary surgical error detection tasks. Code and datasets are available at https://github.com/wzjialang/SEDCLIP.
Helicobacter pylori (H. pylori) infection is a major risk factor for gastric carcinogenesis, requiring holistic endoscopic assessment of distributed mucosal abnormalities. While patient-level diagnostic labels are accessible, obtaining large-scale fine-grained annotations is constrained due to high inter-observer variability and the labor-intensive nature of the process. This naturally formulates the task as a multiple instance learning (MIL) problem. Existing image-level approaches often rely on noisy supervisory signals without modeling sequence context, while standard embedding-based MIL methods suffer from suboptimal feature representations and insufficient integration of H. pylori-specific phenotypes. To address these challenges, we propose PhenoMIL, a two-stage framework emulating clinical hierarchical reasoning. First, the Fine-grained Phenotype-level Semi-Supervised Learning (FPS-SL) stage uses multi-label supervised contrastive learning and prototype-based pseudo-label generation to acquire phenotype-specific knowledge and learn transferable representation from limited instance-level annotations. Subsequently, the Clinically Informed MIL (CIMIL) stage performs holistic bag-level diagnosis using Consensus-aware Feature Modulation (CFM) to exploit latent neighborhood structures, and Phenotype-guided Attention Aggregation (PAA) to align instance weighting with diagnostic priors. Extensive evaluation on a large-scale multi-center dataset (303,910 images from 6388 patients) demonstrates PhenoMIL’s superiority in both multi-label phenotype classification and H. pylori infection diagnosis. Notably, in a real-world reader study, PhenoMIL achieved 81.63% sensitivity and 87.84% specificity, outperforming both junior and senior endoscopists.
Accurate segmentation of tubular structures, such as blood vessels and airway trees, is critical across various medical applications. However, this task remains highly challenging due to the diverse global morphologies and thin local structures of these tubular formations. Segment Anything Model (SAM)-based methods have shown impressive performance in image segmentation, yet their reliance on manually crafted prompts makes them unsuitable for fine-grained tubular structures. Moreover, existing methods for fine-tuning SAM have not fully explored how to identify the most informative parameters to maximize adaptation performance. To address these issues, we introduce a novel framework termed Active Adaptation of SAM (A2SAM), which adopts informativeness-driven adaptation as its governing principle for medical tubular segmentation. Specifically, we design a structural prompt generator that automatically produces topology-aware prompts, guiding SAM toward anatomically meaningful tubular structures. The structural uncertainty embedded in these prompts captures variations in global morphology. Furthermore, we find that parameters with strong task-specific influence and a lower risk of disrupting generalizable knowledge are beneficial for adapting SAM. We define these parameters as contrastive parameters and propose a contrastive parameter selection strategy that modulates squared-gradient influence with magnitude-based preservation. The proposed strategy enhances adaptation performance in tubular segmentation and preserves SAM's generalizable knowledge. Finally, we propose a thickness-aware loss function to improve local connectivity in thin structures. The proposed A2SAM is validated on five diverse benchmark datasets for medical tubular segmentation. Comprehensive experiments demonstrate that A2SAM achieves strong overall performance. Our code is publicly available at: https://github.com/YongChen-Exact/A2SAM.
Automated medical image analysis plays a critical role in disease screening and diagnosis. However, image quality degradation, such as blurring and uneven illumination hinders both clinical interpretation and computer-aided diagnostic performance. Most existing enhancement methods often overlook frequency-domain degradation patterns, leading to over-enhancement or loss of clinically relevant details. To address these limitations, we propose a frequency-aware dual-domain collaborative framework for medical image enhancement, termed FDRNet, comprising two key components in this paper: (1) a frequency-decoupled deblurring module with asymmetric channel integration, which combines global and local views using high- and low-frequency information to preserve fine and broad structural features, and (2) a Retinex-guided illumination compensation module with a multi-scale color preservation unit for accurate estimation and correction. Our framework further incorporates a dual-domain collaboration mechanism into each encoder-decoder block of the deblurring module, enabling joint learning of degradation representations in spatial and frequency domains. Extensive experiments on three medical image modalities, utilizing seven public and clinical datasets demonstrate that our method surpasses both traditional and learning-based enhancement techniques. Evaluations on downstream clinical tasks, including vessel segmentation, polyp segmentation, disease diagnosis, and disease severity grading, further confirm significant improvements in task performance and clinical applicability. Our code is available at: https://github.com/iMED-Lab/FDRNet-PyTorch.
Spine and pelvis diseases impair motor control through pain and neurological dysfunction, resulting in characteristic gait abnormalities. However, existing clinical screening and initial triage rely on static imaging, which overlooks gait patterns and limits efficiency for the primary assessment of these diseases. To enhance efficiency and enable clinically aligned diagnostics, we utilize gait videos and propose a hierarchical diagnostic model for spine and pelvis diseases. Based on gait videos, our method progresses from disease-region localization to pathological characterization, explicitly mimicking the clinical diagnostic process. To capture dynamic gait features, we leverage a Spatiotemporal Transformer that jointly models structural dependencies among body regions and motion rhythms. To ensure diagnostic coherence, we design a clinically inspired Tree Hierarchy Classifier that integrates region localization and pathology assessment, while preventing incompatible cross-level predictions. We collected clinical gait videos from 419 patients to train and validate our proposed model. At the first diagnostic level, the model distinguishes cervical/thoracic from lumbar/pelvis diseases, achieving an accuracy of 76% and an AUC of 85%. The second diagnostic level performs pathological characterization, further classifying cases into tumor and non-tumor subtypes across these regions. We further provide clinical interpretability via spatiotemporal attention visualizations, revealing that the model captures gait patterns consistent with established pathophysiological mechanisms. By aligning gait patterns with the dynamic functional impact of disease, we demonstrate the potential of gait video analysis as an efficient approach for the primary screening of spine and pelvis diseases.
Low-count positron emission tomography (PET) reduces injected activity or acquisition time, but fewer detected coincidence events compromise image quality and quantitative reliability. Existing diffusion-based PET enhancement methods commonly use generic, count-agnostic Gaussian schedules that do not explicitly represent acquisition/count-dependent variation in degradation severity between paired standard- and low-count reconstructions. Direct incorporation of multi-step history may also complicate Markovian reverse inference. We propose a count-informed, endpoint-conditioned diffusion bridge in the reconstructed-image domain with gated autoregressive inference (GAI). The bridge is anchored to paired reconstructed endpoints: its conditional mean follows the standard-to-low-count residual, while a normalized expected-count trajectory controls progression along that residual and the aggregate latent variance. Measurement-level Poisson counting statistics motivate this trajectory from acquisition duration or injected-activity ratio, but no Poisson likelihood is imposed on reconstructed PET voxels. During reverse inference, GAI summarizes previous reverse states within an augmented state that retains a first-order Markov formulation and permits closed-form, count-conditioned updates. We evaluated the method on four-center total-body [18F]FDG short-duration PET datasets and a simulated BrainWeb low-dose dataset. On the Shanghai Ruijin cohort, it increased PSNR by 5.33 dB and SSIM by 0.043 and reduced RMSE by 53.4% for 3s acquisitions relative to the unenhanced short-duration PET (sdPET) input. At 1s, PSNR increased by 5.91 dB and RMSE decreased by more than 60%. BrainWeb provided proof-of-concept evidence under controlled count reduction. These results support count-informed degradation modeling for low-count PET enhancement in the reconstructed-image domain.
Medical image segmentation is crucial for clinical diagnosis and treatment planning. However, in complex scenarios, factors such as boundary ambiguity, annotation noise, and domain shift often lead models to be overconfident, limiting the interpretability and trustworthy deployability. To address this, we propose a "plug-and-play" evidence-guided framework for reliable medical image segmentation. By incorporating an evidential constraint to suppress "high-confidence but erroneous" predictions, and without altering the existing segmentation backbone architecture, we append a lightweight evidential prediction head to enable the model to simultaneously output segmentation results and pixel-wise uncertainty in a single forward pass. An error-driven uncertainty calibration term is further introduced to enhance the joint calibration of evidence strength and scale, improving the stability, interpretability, and consistency of the risk indicator with the true error. Moreover, to reduce the reliance on manual empirical thresholds when utilizing uncertainty, we design an adaptive threshold selection process based on a validation set. During deployment, this enables the output of a reference-free high-risk region prompt for quality control, and, when annotations are available, facilitates quantitative validation through geometric consistency metrics between the uncertain region and the actual error region. We systematically evaluate our framework on a representative backbone, U-Net, across multiple datasets, and further investigate its architectural applicability through cross-backbone experiments. The results demonstrate that, with negligible increase in inference cost, our method generally achieves a more consistent, spatially coherent uncertainty representation that better aligns with the true error, providing a practical tool for quality control and risk awareness in clinical deployment. The source code is available at: https://github.com/Xingchao003/EviRisk-Seg.
For understanding dynamic biological processes such as the life cycle progression of bacteria at the single-cell level, automatic methods for tracking and state classification are needed. In this work, we propose a unified framework based on Graph Neural Networks for simultaneous bacteria tracking, division detection, and life cycle state classification. In previous work, these tasks were treated separately. With our method, trajectories are represented by a graph, where nodes represent bacteria at different time points of a live-cell microscopy video, and edges represent their interactions over multiple frames. Tracking, division detection, and life cycle state classification are performed simultaneously by classifying graph nodes and edges. For all three tasks, we use visual object features from a large-scale pretrained foundation model. This eliminates the need for separately-trained task-specific CNN encoders as used in previous work and enhances the robustness. In addition, we introduce a network-based approach for segmentation error correction using division and multi-frame correspondence predictions. Our method was evaluated using live-cell bright-field microscopy videos of spore germination and outgrowth of rod-shaped bacteria. Our experiments show that the proposed method outperforms existing methods for division detection and state classification. The method yields state-of-the-art results for bacteria tracking and shows increased robustness against segmentation errors as well as image distortions.
