The Oncotype DX® Recurrence Score (RS) is essential for guiding adjuvant chemotherapy decisions in early-stage ER+/HER2- breast cancer, but it is costly, centralized, and not always accessible. This study developed a radiopathomic approach to estimate RS from routinely available data. A radiomic model trained on MRI using LightGBM achieved an AUC of 0.74 on an external test set, while the pathomic model using ElasticNet reached an AUC of 0.74 in cross-validation. While multimodal fusion improved discrimination in 20 paired cases (AUC $=0.74$; balanced accuracy $=0.77$). Integration with Ki67 and PGR within a nomogram further enhanced prediction (AUC $=0.76$). High specificity and NPV (0.93 and 0.88, respectively) supported the reliable identification of lowrisk patients, potentially helping reduce unnecessary chemotherapy. To our knowledge, this represents the first integrated radiopathomic signature for Oncotype DX® prediction, highlighting the feasibility of noninvasive genomic risk estimation and the potential of multiomic AI to support personalized treatment guidance in ER+/HER2breast cancer.
Automated radiology report generation requires complex compositional reasoning to translate subtle visual findings into coherent diagnostic narratives. Current models often fail to capture fine-grained pathological details, leading to generic or factually inconsistent reports. To address these limitations, we propose an expert model guided visual programming framework. It decomposes radiological reasoning into specialized subtasks. Each subtask is executed by a domain-specific expert model. We use a two-stage progressive fine-tuning strategy to fine-tune the model. This architecture enables the vision-language model to leverage the precision of specialized medical vision tools while preserving strong language generation capabilities. We show that small, task-specific models can effectively guide the training of large medical language models. Through our experiments, we show an improved state-of-the-art performance on IU-XRay and MIMIC-CXR datasets.
Deep learning has greatly advanced electron microscopy (EM) segmentation, yet supervised models remain limited by costly voxel annotations and poor generalization across datasets with large domain shifts in contrast, texture, and morphology. These discrepancies distort both appearance and organelle topology, making structure-preserving adaptation essential for reliable EM analysis. To address this challenge, we propose DiSMemNet, a unified framework for unsupervised domain adaptation (UDA) that jointly enhances appearance-structure alignment and cross-domain generalization. The framework consists of two synergistic modules: a gated contrastive translation module that preserves morphological continuity during adversarial alignment, and a distillation-memory synergy module that stabilizes pseudo-label learning via multi-scale distillation and prototype-guided refinement. Through unified optimization of semantic, structural, and appearance consistency, DiSMemNet achieves topology-preserving adaptation and stable convergence across diverse EM domains. Extensive experiments on MitoADWT, SynADWT, and Lucchi datasets demonstrate that DiSMemNet achieves state-of-the-art performance, yielding sharper boundaries, continuous structures, and strong generalization without target annotations. The results highlight its potential as a structure-aware and domainrobust framework for large-scale EM segmentation. Our code is available at https://github.com/RNZhou0529/DiSMemNet.
This paper introduces GLUS, a novel training-free plug-and-play framework for ultrasound (US) super-resolution. GLUS extends the residual plug-and-play paradigm by incorporating a tailored guidance energy function to capture the domain-specific characteristics of US data. The proposed framework operates without the need for pre-training, offering a practical and efficient solution for enhancing US image resolution. Experiments on both simulated and in vivo data show that GLUS consistently outperforms existing benchmarks, achieving superior quantitative and qualitative results.
Personalized Federated Learning (PFL) presents a promising paradigm for mitigating privacy concerns and addressing data heterogeneity in medical images. However, the presence of data heterogeneity can introduce bias into the global model. As a result, the biased global model may degrade local performance in the personalization phase. To tackle this challenge, we propose FedLP, a novel framework that enhances global aggregation fairness and improves personalized performance. First, we propose a layer-wise aggregation strategy on the server, where it dynamically estimates each client's contribution based on the gradient discrepancies. Consider the aggregated global model may not consistently align with local distributions, we then introduce an adaptive parameter personalization process on the client side. By selectively instituting local parameters with the global model parameter with maximizing consensus optimization, clients can better retain global knowledge they need and maintain important model parameters to local data distributions. Experimental results demonstrate that FedLP outperforms state-of-the-art methods and achieves fairer performance for several medical image classification tasks.
The paper proposes a deep learning model for dynamics segmentation and classification of particle trajectories, based only on trajectory points. It enables the classification of the six main stochastic behaviors observed at the cellular membrane, detecting related switch times along trajectories. Compared to existing methods, this highlights a larger family of motion describing the main biophysical constraints at the subcellular scale. Based on an architecture variation of the model proposed in [1], our approach improves model compactness and segmentation accuracy. An application is shown to the effect of an HIV-antiviral compound based on the dynamical analysis of cellular receptors (CCR5).
Reliable decoding of motor imagery (MI) from electroencephalography (EEG) signals remains a core challenge in brain—computer interface (BCI) research, owing to their low signal-to-noise ratio and intricate spatial—temporal dynamics. To address these issues, we propose ST-Mamba, a unified Spatial—Temporal Attentional Convolutional Network with a Dual-Branch BiMamba Encoder for efficient EEG representation learning. The proposed Attention-Guided Spatial—Temporal (AGST) module leverages hierarchical convolution and grouped attention to extract discriminative multiscale features, yielding compact spatio-temporal embeddings. Building on recent advances in state-space modeling, the BiMamba encoder employs bidirectional selective state updates to capture long-range temporal continuity and interchannel correlations with linear computational complexity. A lightweight Temporal Convolutional Network (TCN) further refines sequential dynamics to enhance decoding stability. Comprehensive evaluation on the BCI Competition IV-2a dataset demonstrates ST-Mamba's superior performance in both within-subject and cross-subject settings. These results validate the efficacy of combining state-space modeling with attention mechanisms for robust EEG-based MI decoding.
Mild cognitive impairment (MCI) diagnosis from neuroimaging requires a comprehensive understanding of brain connectivity patterns. While multi-atlas approaches provide complementary perspectives on brain organization, existing methods rely predominantly on functional connectivity and employ static fusion strategies that introduce information redundancy. To address this, we propose a novel framework that integrates individualized three-hinge gyrus (3 HG) -based brain networks with population-level atlases based networks through a Mixture-of-Experts (MoE) architecture. Specifically, brain graph encoders extract structural and functional connectivity embeddings from each atlas, which are processed by atlas-shared and atlas-specific experts to learn atlas-shared and atlas-specific brain patterns, respectively. A gating mechanism dynamically weighs expert contributions, enabling adaptive cross-atlas fusion. Evaluations on the ADNI demonstrate that our method achieves superior performance while providing interpretable insights into atlas-specific diagnostic contributions. Code is available at: https://github.com/Nancy-Zhang-0/Multi-Atlas-MoE
Electroencephalography (EEG) offers a portable, affordable acquisition for BCI applications. However, advanced decoding tasks such as EEG-to-image generation faces significant challenges due to the inherent noisy nature of EEG signals and the availability of limited paired EEG-Image datasets. In this paper we propose an adversarial framework that transforms EEG manifold space to image manifold space for EEG-to-image generation. Our method employs Wasserstein distance-guided feature translation method, enabling effective learning with limited data samples. A multi-view encoder processes both temporal and spectrotemporal EEG representations Through adversarial feature translation to pretrained VQ latent space. To mitigate the codebook collapse that may happen during end-to-end translation, we use MS3D regularization and compute multi-scale self-dissimilarity score to regularize the discriminator. We achieve efficient training and inference suitable for practical BCI applications. Extensive experiments demonstrate state-of-the-art performance: Inception Score of 33.39 and FID of 86.53, substantially outperforming existing GAN-based methods. Our training approach shows competetive results comparable to diffusion-based methods, which require more resources and computation compared to the lightweight proposed model. Our approach establishes adversarial based feature alignment method as an effective paradigm for neuralvisual translation under resource constraints, advancing practical EEG-based image generation for BCI applications. Our source code is available at: https://github.com/29rabhishek/MultiView-EEGWGAN.
Three-dimensional fluorescence microscopy images often suffer from blur and noise induced by the point spread function (PSF), leading to loss of resolution and structural fidelity. We propose EnergyFlow, a method for 3D microscopy deconvolution based on energy-guided flow matching. The model learns a restoration flow field along the continuous physical trajectory of degradation, while an energy consistency constraint enforces physical interpretability throughout the degradation-recovery process. This formulation enables image reconstruction without explicit inversion and achieves superior performance over classical deconvolution and deep learning baselines on both synthetic and real fluorescence data. The results demonstrate that energy-guided flow matching provides a robust and physically consistent framework for 3D microscopy restoration.
Three-dimensional (3D) echocardiography is an inherently noisy modality, but allows for capture of temporally complex cardiac motions. Existing segmentation methods often treat cardiac frames independently, resulting in inconsistent delineations across the cardiac cycle, reduced segmentation accuracy, and, in turn, impacting the estimation of key clinical metrics. To address this, our work introduces temporal positional embeddings based on the cardiac cycle to improve 3D echocardiography segmentation. By encoding cardiac phase information into sinusoidal functions, we inject temporal embeddings into the bottleneck and decoder of a U-Net. Our approach (TU-Net) outperforms state-of-the-art models, including nnU-Net and Transformer baselines, UNETR and SwinUNETR. TU-Net achieved a Dice score of 84.7% along with improved temporal consistency, evaluated over two test cases with 18 and 16 frames across the cardiac cycle. Testing this on a lightweight U-Net, which is both efficient and suitable for clinical settings, demonstrates the significance of temporal information in enhancing segmentation quality without complex models. These results highlight the potential for further improvements, not only in 3D echocardiography but also in other dynamic medical imaging modalities.
This paper presents a nonlinear beamforming technique, the Adaptive Filtered Delay Gaussian-Weighted Multiply-andSum ($a$ FDG $w$ MAS) beamformer, developed for pulsed laser diode (PLD)-based photoacoustic imaging (PAI) systems, where conventional Delay-and-Sum (DAS) and Filtered Delay Multiply-and-Sum (FDMAS) methods suffer from reduced contrast and poor target detection. The $a$ FDG $w$ MAS employs a two-stage adaptive weighting strategy combining Gaussian spatial weighting with data-driven adaptive filtering to enhance signal coherence and suppress background noise. Experimental validation using chicken breast tissue containing three Indian ink-filled capillary targets demonstrates significant improvements over DAS and FDMAS. Quantitatively, the method achieves contrast ratio (CR) and contrast-to-noise ratio (CNR) enhancements of up to 426 % and 153 % at $28 \mu ~\mathrm{J}$, and 263 % and 180 % at $21 \mu ~\mathrm{J}$, respectively. Lateral profile analysis further confirms improved target localization, reduced sidelobes, and enhanced contrast, validating the robustness of $a$ FDG $w$ MAS for compact, low-energy PLD-based PAI systems.
Intracellular particles like viruses, synaptic or transport vesicles are known to switch between Brownian motion and directed motion to reach their destination (e.g nucleus or cell membrane). We model this situation with a mixture of Brownian and Brownian with drift (or directed Brownian). Hence particles movements are modeled with a finite number of diffusion coefficients and velocity drifts. For more flexibility, velocity drifts are constrained by their norm but not their direction. An original EM algorithm is proposed to estimate these parameters. The estimation process can deal easily with thousands of long or very short trajectories in a matter of seconds.
Ischemic stroke remains a leading cause of death worldwide. Ultrasound imaging plays a vital role in evaluating carotid plaques, but automatic segmentation and vulnerability assessment face significant challenges due to high annotation costs and subjective manual evaluation. To address these issues, this paper proposes a semi-supervised learning framework that leverages abundant unlabeled ultrasound images to reduce reliance on costly labeled data. Our approach introduces three key improvements: a cascaded dual-model architecture decoupling segmentation and classification tasks; an enhanced semi-supervised training strategy with optimal teacher model and dynamic threshold pseudo-label filtering; and a differentiated loss function to mitigate class imbalance. These innovations significantly reduce manual annotation demands while maintaining high accuracy. Experimental validation on the CSV2026 dataset demonstrates the effectiveness of our approach, ultimately securing third place on the final test set leaderboard and providing reliable support for clinical stroke risk evaluation.
Histopathological image segmentation is crucial for clinical diagnosis, but obtaining high-quality annotations is timeconsuming and labor-intensive. To improve segmentation performance while reducing annotation dependence, this paper presents DisDiff, a novel two-stage self-supervised framework leveraging diffusion models for nuclei segmentation without pseudo-labels. In the first stage, a Denoising Diffusion Probabilistic Model (DDPM) with Perception Prioritized (P2) weighting learns robust features from unlabeled images. In the second stage, pretrained weights are transferred to a U-Net enhanced with Efficient Multi-scale Attention (EMA) and distance map post-processing for effective nuclei separation. Experiments on TCGA and TNBC datasets show that DisDiff achieves competitive performance with only 50 % labeled data and superior results with full annotations, demonstrating practical value for clinical applications.
MRI is highly sensitive to site-specific acquisition conditions (e.g., scanner models, protocols). To avoid unwanted biases, multicentric studies require a harmonization step to mitigate these scanner-specific effects. One widely used method is ComBat, which has proven effective when a sufficiently large number of subjects per site, preferably healthy controls (HC), are available to ensure reliable statistical estimation. In clinical practice, however, sites want to scan very few HC subjects, making a blind application of ComBat unreliable and potentially leading to erroneous harmonization. We introduce Frugal-ComBat, a Bayesian method that leverages well-populated third-party sites to augment a given site's population by integrating information from a broader community. Experimental results demonstrate that with as few as 1 or 2 HC subjects, Frugal-ComBat effectively achieves up to 90% correct harmonizations for HC, with consistently lower errors compared to both the absence of harmonization and ComBat.
Multi-view light-sheet microscopy captures complementary information from different angles to obtain a comprehensive representation of the specimen. However, fusing these views into a single, high-fidelity reconstruction is challenging, particularly under significant noise. Traditional methods like content-weighted fusion and deconvolution struggle with noise, while supervised deep learning requires difficult-to-acquire, high-quality ground-truth data. To circumvent these limitations, we propose Guided Implicit Fusion (GIF), a self-supervised framework that represents the specimen as a continuous function using an Implicit Neural Representation. By leveraging the image formation model as a constraint, GIF directly optimizes the object representation from the low-SNR multiview data itself, obviating the need for external labels. It enables fusion, denoising, and deconvolution within a unified framework. Evaluations on simulated and experimental datasets show GIF provides high-quality, efficient multi-view fusion, outperforming both traditional and supervised methods, especially under noisy conditions.
In endoscopic imaging, monocular depth estimation (MDE) is crucial for enhancing 3D scene understanding, tissue assessment, and improving diagnostic and surgical precision. However, depth estimation around edges remains a challenge, as standard regression approaches often produce overly smooth predictions and blurred discontinuities. To address this, we propose a framework that integrates edge-aware supervision with a guidance-driven decoder to achieve precise edge preservation and structure-consistent depth reconstruction. First, we introduce EdgeAttenLoss, an edge-aware loss that employs attention maps to emphasise edge regions during training, enhancing boundary precision. Second, a lightweight Guidance Map Extractor encodes the RGB input into spatially rich features that provide structural cues for decoding. Finally, a Guided-Affine Decoder fuses encoder features with the learned guidance via spatially adaptive affine modulation, producing high-quality, detail-preserving depth maps. Evaluations on public ‘UCL’ and ‘C3VD’ endoscopic datasets demonstrate the effectiveness of our depth estimation approach around edges and detailed structures.
Vision loss remains a critical global health issue, with conditions such as diabetic macular edema and age-related macular degeneration often detected only after substantial retinal damage. Early retinal fluid accumulation is a key biomarker, and Optical Coherence Tomography (OCT) provides noninvasive, high-resolution retinal imaging essential for diagnosis. Accurate segmentation of retinal structures is crucial, yet recent conventional deep learning approaches often do not provide satisfactory performance because fluid regions occupy a small portion of the image, leading to inefficient computation and reduced accuracy. We propose CLAVIS, a novel plug-and-play pipeline that leverages CLAssification outputs to generate VIsualizations and guide Segmentation, generating more precise maps of the fluid regions. We leverage visualization techniques to target image regions that influenced the classification outcome. Experimental results demonstrate that this guidance strategy improves Dice score by $\mathbf{+ 0. 1 4}$ on the average and up to $\mathbf{+ 0. 1 8}$ for some pathologies.
Acquiring high-quality synthetic computed tomography (sCT) images from cone-beam CT is crucial for accurate dose calculation in the field of adaptive radiotherapy. However, existing sCT generation methods based on diffusion models often struggle to preserve anatomical structures, especially when the anatomical structure of the input CBCT and ground-truth CT differs in shape. To address this, we propose a Hierarchical Anatomical-Guided Diffusion (HAGD) framework that captures both global semantic information and local textural details for anatomical-aware sCT generation. We first introduce a dual-branch encoder to simultaneously extract features from coarse structural masks and fine segmentation maps. Secondly, a hierarchical fusion module is proposed to integrate these multiscale features into the conditional diffusion model, using concatenation or attention-gated fusion. Experiments on the SynthRAD2023 dataset show that HAGD achieves superior performance, with a mean absolute error (MAE) of 56.51 for concatenation fusion and 54.62 for attention-gated fusion improves performance over existing image translation models. These results highlight the effectiveness of structured anatomical guidance in enhancing fidelity and structural consistency of sCT images, thus supporting accurate and robust image translation in clinical radiotherapy workflows.
