Federated learning (FL) enables collaborative medical image classification without sharing raw data, but its performance is often limited by negative transfer arising from heterogeneous client distributions and privacy constraints. Existing methods, including federated distillation and data-driven domain generalization, either lack explicit mechanisms to address inter-client discrepancies or require additional information sharing. In this paper, we propose an Interim-driven Hierarchical Alignment Federated Network (IHAFN) to address the generalization problem in heterogeneous settings. IHAFN performs progressive alignment from feature-level to prototype-level. In the first stage, a Local-Interim Collaborative Optimization Strategy (LICOS), together with an interim model, reduces representation discrepancies across clients. A Bidirectional Divergence loss is introduced to enforce consistent yet discriminative feature alignment. In the second stage, a Prototype-Based Alignment Strategy (PBAS) promotes domainagnostic feature representations and improves generalization to unseen domains. Extensive experiments on breast cancer and prostate cancer classification tasks demonstrate that IHAFN consistently outperforms state-of-the-art federated learning and domain generalization methods under heterogeneous and privacy-constrained scenarios.

DaHAC combines dual large language models with human expert adjudication to extract high-fidelity CRISPR biosensing knowledge from literature, building a validated dataset and benchmark for evaluating LLM reasoning in biosensing.

IntroductionThis study aims to develop a noninvasive CT-based radiogenomic framework to estimate imaging-associated biomarkers related to natural killer (NK) cell cytotoxicity in clear cell renal cell carcinoma (ccRCC). By linking imaging features to underlying molecular pathways, we address the limitations of invasive tissue sampling in assessing tumor heterogeneity and immune microenvironment.MethodsThis study analyzed preoperative contrast-enhanced CT images from 143 patients with histologically confirmed ccRCC and transcriptomic data from 538 TCGA-KIRC tumor samples. Radiomic features (3,176 per tumor) were extracted using PyRadiomics from manually segmented 3D tumor volumes (validated by two radiologists), spanning seven feature classes and eight image filters. Feature selection included variance filtering, the Mann-Whitney U test for gene-expression stratification (high/low groups), and redundancy removal via Pearson correlation analysis (|r| > 0.9), with AUC prioritization. An L1-penalized support vector machine model with leave-one-out cross-validation was developed to estimate expression levels of ICAM1 and RAET1E, two NK cytotoxicity-related biomarkers. External validation was performed using tissue microarrays and immunohistochemistry from 26 independent ccRCC cases. Differential gene expression was assessed using edgeR (FDR < 0.05, fold-change > 2) and survival analysis via Cox regression.ResultsTranscriptomic analysis identified 835 imaging-associated genes enriched in immune-related pathways, with significant representation of the NK cell-mediated cytotoxicity pathway (KEGG, P < 0.05). Among candidate genes, ICAM1 and RAET1E demonstrated the strongest radiogenomic associations (AUC = 75.7% and 67.3%, respectively). Immunohistochemistry confirmed increased ICAM1 and decreased RAET1E expression in ccRCC tissues. Higher ICAM1 and lower RAET1E expression were associated with advanced tumor stage and unfavorable survival outcomes. External validation of the L1-SVM model achieved predictive accuracies of 76.92% for ICAM1 and 73.08% for RAET1E, supporting the preliminary feasibility of this radiogenomic approach.ConclusionsOur findings suggest that CT-derived radiomic features may provide noninvasive imaging correlates of biomarkers related to NK cytotoxicity-in ccRCC. Radiogenomic analysis of ICAM1 and RAET1E may provide a complementary exploratory framework for noninvasive immune characterization and biomarker research in ccRCC.

Federated learning (FL) has emerged as a pivotal paradigm for medical image classification in multi-client healthcare applications by enabling collaborative training without sharing raw data. However, its practical deployment remains challenged by negative transfer caused by heterogeneous client distributions and privacy risks introduced by data-sharing-based domain generalization strategies. Existing approaches, including federated model distillation and data-driven domain generalization methods, either lack explicit mechanisms to alleviate inter-client discrepancies or rely on additional information sharing that may compromise privacy. To address these challenges, we propose a Dual-stage Hierarchical Alignment Federated Network (DHA-Fed), which establishes a principled alignment framework from feature-level to prototype-level. In the first stage, a Local-Interim Model Synergistic Feature-Level Alignment strategy (LMSFA) and an interim model mechanism are introduced to explicitly reduce representation discrepancies across clients, thereby mitigating negative transfer. A Bidirectional Divergence loss is further designed to enforce consistent feature alignment while preserving discriminative capability. In the second stage, a Prototype Relaxation Strategy (PRS) is proposed to relax the model's error bound on unseen domains and foster domain-agnostic feature manifolds through prototype contrastive minimization. Extensive experiments on two real-world clinical tasks, including breast cancer and prostate cancer classification, demonstrate that DHA-Fed consistently outperforms state-of-the-art federated learning and domain generalization methods under heterogeneous and privacy-constrained settings.

Due to immunosuppression, mucosal barrier injury, and prolonged neutropenia resulting from both the disease and chemotherapy, along with the frequent use of broad-spectrum antibiotics and glucocorticoids, children with leukemia are at a high risk of invasive fungal disease (IFD). The present study aimed to develop an effective machine learning model to predict fungal infections in children with leukemia. A total of 247 pediatric patients diagnosed with leukemia and concurrent infections were evaluated. Five distinct ML classifiers—Random Forest, Logistic Regression, Support Vector Machine (SVM), Naïve Bayes, and K-Nearest Neighbors—were employed to construct predictive classification models. These models were trained using three distinct feature sets: (1) clinical features exclusively, (2) imaging features exclusively, and (3) an integrated feature set comprising both clinical and imaging data. The predictive model was validated prospectively in an independent cohort of 61 patients. Model performance was evaluated through cross-validation techniques to ensure robustness and generalizability. To validate the clinical applicability of the ML models, their diagnostic performance was systematically compared against that of three radiologists with varying experience levels: Reader A (3 years), Reader B (6 years), and Reader C (11 years). Among the five classifiers evaluated, models using both clinical and imaging features consistently outperformed those relying solely on either clinical or imaging features. Notably, the SVM algorithm exhibited the highest overall predictive performance. Within the SVM algorithm, the validation set achieved the mean area under the curve (AUC) values of 0.825 with clinical features alone, 0.852 with imaging features alone, and 0.947 when both clinical and imaging features were combined. The corresponding mean AUC values for the test set were 0.777, 0.797, and 0.879. Furthermore, a comparative analysis between the classification results of the SVM model and the diagnostic assessments provided by three radiologists demonstrated that the SVM consistently outperformed the radiologists across key performance metrics. The SVM algorithm demonstrates robust efficacy in predicting fungal infections among pediatric patients diagnosed with leukemia. Within the predictive model, the variables that exhibited the greatest influence included pleural thickening, neutropenia, hormone therapy, CRP level, mediastinal lymphadenopathy, and the presence of pleural effusion.

This study introduces an innovative method for early ultrasound classification of developmental dysplasia of the hip (DDH) in infants, integrating key point detection based on the YOLOv8 model with radiomics. Leveraging YOLOv8's advanced capabilities, the method accurately identifies and locates key anatomical landmarks in pediatric hip ultrasound images. Constructing angular information through these key points serves as a crucial clinical feature for DDH classification. By fusing these clinical features with radiomic features, the method further characterizes hip joint morphology and function. The aim is to enhance the accuracy and reliability of early DDH classification, facilitating timely intervention and improving patient prognosis. Preliminary results demonstrate the feasibility and potential of this integrated approach, with an average α Angle Difference of 3.13 degrees and an average β Angle Difference of 4.83 degrees in hip DDH imaging. The model's automatic calculation of α and β angles is comparable to clinicians, alleviating clinical burden while providing valuable clinical features for DDH classification. Furthermore, based on deep learning analysis combining clinical and radiomic features, the model has achieved impressive results on multi-center test sets, with an accuracy, recall, precision, and F1 score of 0.8544, 0.8321, 0.8465, and 0.8383 respectively.

Vascular segmentation is essential for various medical applications, such as computer-aided diagnosis, treatment planning, and surgical interventions. However, current deep learning-based vascular segmentation methods face two significant challenges: the complex morphological diversity of vascular structures, which results in discontinuous segmentation in small-scale vessels and incomplete preservation of topological integrity, and the adverse effects of noisy labels during network optimization. To address these challenges, we propose a Joint Hierarchical Vascular Morphology Learning and Noise Label Refinement method (JHN-Seg). JHN-Seg introduces a Hierarchical Vascular Morphology-Aware Network (HVMA-Net) that integrates a Multi-Scale Local Morphology-Aware (MLMA) module, employing a multi-pattern convolutional strategy to adaptively capture intricate vascular features across scales. A Global Morphology Preserving (GMP) loss function is incorporated into HVMA-Net to enforce the continuity of small-scale vessels and maintain the integrality of global vascular structures. Furthermore, JHN-Seg introduces an Uncertainty-Aware Distillation (UD) strategy, which incorporates an Uncertainty Label Refinement (ULR) Module for uncertainty-guided noisy label correction by leveraging pixel-wise KL divergence and consistency generated by teacher-student framework. Comprehensive experiments on liver vessel datasets demonstrate that JHN-Seg outperforms other state-of-the-art segmentation methods. The framework's adaptability and performance advancements position it as a transformative solution for vascular segmentation and broad applicability to medical image analysis tasks requiring precise morphological representation and noise-robust learning.

Segmentation of multiple targets of varying sizes within medical images is of significant importance for the diagnosis of disease and pathological research. Transformer-based methods are emerging in the medical image segmentation, leveraging the powerful yet computationally intensive self-attention mechanism. A variety of attention mechanisms have been proposed to reduce computation at the cost of accuracy loss, utilizing handcrafted patterns within local or artificially defined receptive fields. Furthermore, the common region-based loss functions are insufficient for guiding the transformer to focus on tissue regions, resulting in their unsuitability for the segmentation of tissues with intricate boundaries. This paper presents the development of a bi-level sparse attention network and a narrow band (NB) loss function for the accurate and efficient multi-target segmentation of medical images. In particular, we introduce a bi-level sparse attention module (BSAM) and formulate a segmentation network based on this module. The BSAM consists of coarse-grained patch-level attention and fine-grained pixel-level attention, which captures fine-grained contextual features in adaptive receptive fields learned by patch-level attention. This results in enhanced segmentation accuracy while simultaneously reducing computational complexity. The proposed narrow-band (NB) loss function constructs a target region in close proximity to the tissue boundary. The network is thus guided to perform boundary-aware segmentation, thereby simultaneously alleviating the issues of over-segmentation and under-segmentation. A series of comprehensive experiments on whole brains, brain tumors and abdominal organs, demonstrate that our method outperforms other state-of-the-art segmentation methods. Furthermore, the BSAM and NB loss can be applied flexibly to a variety of network frameworks.

Our work aimed to assess the clinical value of dual-energy computed tomography (DECT) parameters in predicting in-stent restenosis (ISR) after carotid artery stenting (CAS) and to develop a nomogram model incorporating these parameters to enhance the accuracy of ISR risk prediction. Our retrospective multicenter research enrolled 205 patients who underwent CAS from January 2018 to April 2023, with DECT scans performed prior to the procedure. Two radiologists independently measured the DECT parameters and evaluated the characteristics of carotid plaques. Univariate and multivariate analyses were conducted to identify independent predictors of ISR. Three models were developed: clinical model, DECT model, and nomogram model. These models were assessed based on the area under the curve (AUC) and calibration, with their clinical value assessed utilizing decision curve analysis (DCA). Among 205 patients, 35 in the training set and 15 in the validation set experienced ISR. Multivariate analysis identified plaque length, fat fraction, normalized iodine concentration, and effective atomic number as independent predictors of ISR. Nomogram model, combining clinical and DECT parameters, demonstrated high accuracy in predicting ISR, with 0.931 (95

Background:Early alterations in cerebrospinal fluid (CSF) may play a critical role in the progression of acute ischemic stroke (AIS). This study aimed to develop a CSF-based clinical-radiomics model for predicting the functional outcomes in AIS patients following intravenous thrombolysis (IVT). Methods:This study included 308 AIS patients within 6 h of onset, divided into a training set (n=246) and a hold-out test set (n=62). Functional outcome was assessed at discharge using the modified Rankin Scale (mRS), dichotomized into good (mRS ≤2) and poor (mRS >2) outcomes. CSF regions were automatically segmented on non-contrast computed tomography images, and radiomics features were extracted. After feature selection was sequentially performed using minimum redundancy maximum relevance algorithm followed by least absolute shrinkage and selection operator, the radiomics signature model was constructed. Clinical features were selected through univariate and multivariate logistic regressions and subsequently integrated with radiomics features to develop combined models. Three machine learning classifiers, including Nu Support Vector Classification (NuSVC), logistic regression, and random forest, were trained and tested for prognostic prediction. Model performance was evaluated using receiver operating characteristic curve analysis, decision curve analysis (DCA), and calibration curves, with internal validation via five-fold cross-validation. Results:Among the 308 patients included, 155 patients had a good outcome and 153 had a poor outcome. A total of 1,874 radiomics features were extracted, among which 21 were selected for the radiomics model. Three clinical features were also identified for the clinical model. The combined clinical-radiomics model significantly outperformed single-modality models across all classifiers. The NuSVC-based combined model achieved the best performance, with an area under the curve (AUC) of 0.870 [95% confidence interval (CI): 0.850-0.889] in cross-validation and 0.893 (95% CI: 0.817-0.968) in the test cohort. DCA further confirmed the superior clinical utility of the combined model. Conclusions:This study demonstrates the value of CSF radiomics signatures in predicting functional outcomes in AIS patients after IVT. The CSF-based combined model exhibited strong prognostic performance and clinical utility, highlighting its potential for supporting treatment decision-making.

Glioma segmentation is crucial for brain tumor diagnosis, surgical planning, and prognosis assessment. Current glioma segmentation methods lack effective integration of 3D spatial and 2D multi-view features, especially in heterogeneous MRI data with varying tumor patterns across views and indistinct boundaries. We propose a boundary enhanced multi-view detail capture network, BMVDC-Net, using 3D-UNet architecture that effectively captures multi-view tumor distribution patterns and precise boundaries. First, we propose a multi-view feature representation module consisting of multi-scale axial, sagittal, coronal, and 3D convolutions to extract multidirectional tumor shape details. Second, we propose a boundary enhancement module that employs encoderdecoder parameter sharing, random boundary expansion, and attention fusion to make the network focus on tumor boundaries. Finally, we use post-processing to further refine the segmentation output. Experimental results on BraTS2019 and BraTS2023-GLI datasets demonstrate superior segmentation performance with 0.9 to 6.6 percent points improvement in mean Dice coefficient compared to state-of-the-art methods.

This study aims to create a radiomics nomogram using dual-energy computed tomography (DECT) virtual monoenergetic images (VMI) to accurately identify symptomatic carotid plaques. Between January 2018 and May 2023, data from 416 patients were collected from two centers for retrospective analysis. Center 1 provided data for the training (n = 213) and internal validation (n = 93) sets, and center 2 supplied the external validation set (n = 110). Plaques imaged at 40 keV, 70 keV, and 100 keV were outlined, and the selected radiomics features were used to establish the radiomics model. The classifier with the highest area under the curve (AUC) in the training set generated the radiomics score (Rad-Score). Logistic regression was used to identify risk factors and establish a clinical model. A radiomics nomogram integrating the Rad-score and clinical risk factors was constructed. The predictive performance was evaluated using receiver operating characteristic (ROC) analysis and decision curve analysis (DCA). Plaque ulceration and plaque burden are independent risk factors for symptomatic carotid plaques. The 40 + 70 keV radiomics model achieved excellent diagnostic performance, with an average AUC of 0.805 across all validation sets. Furthermore, the radiomics nomogram, integrating the Rad-score with clinical predictors, demonstrated robust diagnostic accuracy, with AUCs of 0.909, 0.850, and 0.804 in the training, internal validation, and external validation sets, respectively. DCA results suggested that the nomogram was clinically valuable. Our study developed and validated a DECT VMI-based radiomics nomogram for early identification of symptomatic carotid plaques, which can be used to assist clinical diagnosis and treatment decisions. The study introduces an innovative radiomics nomogram utilizing DECT VMI to discern symptomatic carotid plaques with high precision.

To construct an artificial intelligence (AI)-assisted model for identifying the infraorbital posterior ethmoid cells (IPECs) based on deep learning using sagittal CT images. Sagittal CT images of 277 samples with and 142 samples without IPECs were retrospectively collected. An experienced radiologist engaged in the relevant aspects picked a sagittal CT image that best showed IPECs. The images were randomly assigned to the training and test sets, with 541 sides in the training set and 97 sides in the test set. The training set was used to perform a five-fold cross-validation, and the results of each fold were used to predict the test set. The model was built using nnUNet, and its performance was evaluated using Dice and standard classification metrics. The model achieved a Dice coefficient of 0.900 in the training set and 0.891 in the additional set. Precision was 0.965 for the training set and 1.000 for the additional set, while sensitivity was 0.981 and 0.967, respectively. A comparison of the diagnostic efficacy between manual outlining by a less-experienced radiologist and AI-assisted outlining showed a significant improvement in detection efficiency (P < 0.05). The AI model aided correctly in identifying and outlining all IPECs, including 12 sides that the radiologist should improve portraying. AI models can help radiologists identify the IPECs, which can further prompt relevant clinical interventions.

ObjectivesTo explore the utility of gray-scale ultrasound (GSUS) and mammography (MG) for radiomic analysis in distinguishing between breast adenosis and invasive ductal carcinoma (IDC).MethodsData from 147 female patients with pathologically confirmed breast lesions (breast adenosis: 61 patients; IDC: 86 patients) between January 2018 and December 2022 were retrospectively collected. A training cohort of 113 patients (breast adenosis: 50 patients; IDC: 63 patients) diagnosed from January 2018 to December 2021 and a time-independent test cohort of 34 patients (breast adenosis: 11 patients; IDC: 23 patients) diagnosed from January 2022 to December 2022 were included. Radiomic features of lesions were extracted from MG and GSUS images. The least absolute shrinkage and selection operator (LASSO) regression was applied to select the most discriminant features, followed by logistic regression (LR) to construct clinical and radiomic models, as well as a combined model merging radiomic and clinical features. Model performance was assessed using receiver operating characteristic (ROC) analysis.ResultsIn the training cohort, the area under the curve (AUC) for radiomic models based on MG features, GSUS features, and their combination were 0.974, 0.936, and 0.991, respectively. In the test cohort, the AUCs were 0.885, 0.876, and 0.949, respectively. The combined model, incorporating clinical and all radiomic features, and the MG plus GSUS radiomics model were found to exhibit significantly higher AUCs than the clinical model in both the training cohort and test cohort (p<0.05). No significant differences were observed between the combined model and the MG plus GSUS radiomics model in the training cohort and test cohort (p>0.05).ConclusionThe effectiveness of radiomic features derived from GSUS and MG in distinguishing between breast adenosis and IDC is demonstrated. Superior discriminatory efficacy is shown by the combined model, integrating both modalities.

The assessment of deformable registration uncertainty is an important task for the safety and reliability of registration methods in clinical applications. However, it is typically done by a manual and time-consuming procedure. We propose a novel automatic method to predict registration uncertainty based on multi-category features and supervised learning. Three types of features, including deformation field statistical features, deformation field physiologically realistic features, and image similarity features, are introduced and calculated to train the random forest regressor for local registration uncertain prediction. Deformation field statistical features represent the numerical stability of registration optimization, which are correlated to the uncertainty of deformation fields; deformation field physiologically realistic features represent the biomechanical properties of organ motions, which mathematically reflect the physiological reality of deformation; image similarity features reflect the similarity between the warped image and fixed image. The multi-category features comprehensively reflect the registration uncertainty. The strategy of spatial adaptive random perturbations is also introduced to accurately simulate spatial distribution of registration uncertainty, which makes deformation field statistical features more discriminative to the uncertainty of deformation fields. Experiments were conducted on three publicly available thoracic CT image datasets. Seventeen randomly selected image pairs are used to train the random forest model, and 9 image pairs are used to evaluate the prediction model. The quantitative experiments on lung CT images show that the proposed method outperforms the baseline method for uncertain prediction of classical iterative optimization-based registration and deep learning-based registration with different registration qualities. The proposed method achieves good performance for registration uncertain prediction, which has great potential in improving the accuracy of registration uncertain prediction.

In the past few years, convolutional neural networks (CNNs) have been a major focus in medical image registration. However, it has been proved that CNNs are limited in their ability to represent modal-independent feature and understand the spatial correspondence between different modalities. Therefore, we present CBCRnet for the effective feature representation and correspondence. 1) We propose a novel contrast-reconstruction tasks guided pretraining method for modal-independent feature learning and the unaligned image pairs can be directly imported for pretraining. 2) We propose a bidirectional cross modal attention module to capture the explicit spatial correspondence.Clinical Relevance— Multi-modal deformable medical image registration has many applications in diagnostic medical imaging, organ mapping and surgical navigation [1], such as ablation surgery guided by intraprocedural CT and preoperative MR. Therefore, multi-modal deformable image registration is important in its clinical applications.

Developing an efficient embedded feature selection method for both binary and multiclass classification problems is a fundamental topic to be further studied. In this paper, the conventional radial basis function (RBF) network is modified for joint feature selection and classification. By utilizing the anisotropic Gaussian basis function, an individual weight parameter is assigned to each feature for feature weighting. The L1 1regularized loss function is designed by combining the softmax cross entropy loss with L1 1 regularization terms imposing on feature and output weights. A specialized active-set limited-memory projected Quasi-Newton (SAL-PQN) algorithm is proposed to minimize the L1 1-regularized loss function, which can optimize all the adjustable parameters and zero out the redundant feature and output weights simultaneously. Furthermore, a safe dynamic pruning strategy is integrated into the SAL-PQN algorithm for dynamic sparse training of the modified RBF network, which continuously excludes the dynamically zeroed-out feature weights from the training process. It is theoretically assured of yielding the same gradient updates as SAL-PQN, thus the relevant computation of the zeroed-out feature weights in the forward and backward propagations can be safely eliminated. The experimental results demonstrate the effectiveness and efficiency of the proposed methods for joint feature selection and classification, and illustrate the practical utility of the inherent feature weighting capability.

PURPOSE:Multimodal registration is a key task in medical image analysis. Due to the large differences of multimodal images in intensity scale and texture pattern, it is a great challenge to design distinctive similarity metrics to guide deep learning-based multimodal image registration. Besides, since the limitation of the small receptive field, existing deep learning-based methods are mainly suitable for small deformation, but helpless for large deformation. To address the above issues, we present an unsupervised multimodal image registration method based on the multiscale integrated spatial-weight module and dual similarity guidance. METHODS:In this method, a U-shape network with our multiscale integrated spatial-weight module is embedded into a multi-resolution image registration architecture to achieve end-to-end large deformation registration, where the spatial-weight module can effectively highlight the regions with large deformation and aggregate discriminative features, and the multi-resolution architecture further helps to solve the optimization problem of the network in a coarse-to-fine pattern. Furthermore, we introduce a special loss function based on dual similarity, which represents both global gray-scale similarity and local feature similarity, to optimize the unsupervised multimodal registration network. RESULTS:We verified the effectiveness of the proposed method on liver CT-MR images. Experimental results indicate that the proposed method achieves the optimal DSC value and TRE value of 92.70 ± 1.75(%) and 6.52 ± 2.94(mm), compared with other state-of-the-art registration algorithms. CONCLUSION:The proposed method can accurately estimate the large deformation field by aggregating multiscale features, and achieve higher registration accuracy and fast registration speed. Comparative experiments also demonstrate the effectiveness and generalization ability of the algorithm.