Alzheimer's disease (AD) detection from spontaneous speech provides a noninvasive and scalable alternative to traditional diagnostics but remains challenged by limited labeled data in target languages. This paper presents a modular multimodal method for robust AD detection in label-scarce languages without language-specific supervision. The approach integrates three complementary feature streams-acoustic, paralinguistic, and semantic-through dedicated modules: ARFU for cognitively salient feature compression, DIFA for directed intra-modal refinement, and MASA for task-adaptive modality integration. On the ADReSS-M benchmark, our method achieves 91.30% accuracy and 90.91% F1 on the Greek test set, surpassing the previous best by 4.61% and maintaining strong robustness in zero- and few-shot settings. These results confirm the effectiveness of structured fusion and language-agnostic modeling for reliable AD detection in label-scarce language scenarios.

We present DuST-Net, a dual-stream temporal network for early Alzheimer's disease screening from spontaneous speech. The model explicitly decouples and jointly models microlevel intra-sentential dynamics and macro-level inter-sentential discourse flow via two parallel streams, which together ingest a four-component feature set: paralinguistic cues, global discourse descriptors, and sentence-level acoustic and linguistic features. Finally, DuST-Net is trained and evaluated on the PROCESS dataset. The results show that our model outperforms all baselines and significantly improves the Macro-F1 Score. Additional analyses demonstrate that differentiating between temporal scales and modalities is crucial for achieving balanced and robust AD screening from speech.

Polyp segmentation plays a critical role in the diagnosis of colorectal cancer but remains challenging due to the diverse morphologies and indistinct boundaries of polyps. These challenges are particularly prominent in small and irregularly shaped polyps, where precise edge detail capture is essential. While the Segment Anything Model 2 (SAM2) has demonstrated remarkable progress in medical image segmentation through its strong pre-training capabilities, its reliance on global feature extraction strategies limits its ability to model local details and fine-grained features, particularly for complex edges. To address these shortcomings, we propose MSA-SAM2Net, an enhanced framework that integrates a Large Kernel Size Attention (LKSA) module and an Edge-Boosting Module (EBM) to improve feature extraction and edge detail perception. Specifically, LKSA employs multi-scale and gating mechanisms to capture both global context and local details, producing richer attention maps. Meanwhile, EBM enhances segmentation accuracy by focusing on edge regions. Additionally, we introduce a Large-Kernel Grouped Attention (LGA) to further refine feature selection and fusion. Comprehensive experiments on five publicly available and challenging polyp segmentation datasets demonstrate the effectiveness of MSA-SAM2Net, achieving state-of-the-art performance on all five datasets. The source code will be available at: https://github.com/coderMelon0216/MSA-SAM2Net

Label quality critically impacts the performance of federated learning.However, client data often contains varying degrees of label noise.To address this, we propose FedDMC, a framework featuring a dual-model dynamic interaction mechanism to mitigate error accumulation in high-noise scenarios.FedDMC integrates multi-stage noise correction, and an adaptive regularization strategy to improve robustness against complex noise distributions.We evaluate FedDMC on benchmark datasets under diverse heterogeneous noise settings and compare it with state-of-the-art methods, demonstrating its superior performance.Furthermore, we validate its applicability in realworld medical imaging tasks using ADNI-MRI and ADNI-PET datasets.Experimental results confirm that FedDMC consistently outperforms existing approaches under complex heterogeneous label noise.

Accurate identification and segmentation of the cochlea is crucial in the diagnosis and treatment of Vestibular Schwannoma (VS). High-resolution T2-weighted (hrT2) imaging is safer, less expensive, and provides more detailed soft tissue diagnostic information than the commonly used contrast-enhanced T1 (ceT1) imaging. However, acquiring pixel-level hrT2 expert annotation data is costly and time-consuming, limiting its data availability. To this end, we propose an unsupervised domain adaptation model, CE-UDA, and design the Contour Enhancement Module (CEM) module for different imaging characteristics of ceT1 and hrT2 to perform structural boundary enhancement and ultimately generate high-quality pseudo-hrT2 images. Experiments validated on the CrossMoDA2022 dataset show that CE-UDA significantly improves the translation quality of pseudo hrT2 images in the ceT1 to hrT2 domain adaptation task, which in turn improves the segmentation performance of the downstream task.

Mild Cognitive Impairment (MCI) is the early stage of Alzheimer’s Disease(AD). Using speech signals to diagnose MCI offers the advantages of low cost and high efficiency. In this paper, we propose a two-stage model for MCI recognition that employs ECAPA-TDNN as the feature extraction network. To address the challenging task of extracting AD-related features from speech, we introduce a one-dimensional Squeeze-Excitation (SE) block in the feature extraction network. This block adaptively reweights the input features to emphasize those that contribute significantly to the task while suppressing irrelevant or redundant ones. we utilize Chinese spontaneous speech from patients as our research data. After polling prediction, our accuracy in recognizing AD and MCI reaches 94.29% and 85.84%, respectively, demonstrating potential clinical applications.

Recommender systems struggle with data sparsity, but recent research suggests that incorporating high-quality side information is promising for improving performance.This paper proposes a novel knowledge-augmented recommender system that merges a user-item interaction graph with a knowledge graph.We explore relevant knowledge connections by attention mechanism, identify user-preferred item attributes through crossitem similarity scores, and use masked autoencoders to emphasize these connections and features further.Additionally, we cut off unreliable connections and ensure signal alignment between the graphs.Our experiments on Last-FM and MIND datasets demonstrate that our work outperforms existing recommender systems.

Federated learning has demonstrated applicability in extracting insights from homogeneous data across clients.However, in real-world scenarios, the prevalence of multimodal data is on the rise, driven by the complex nature of industrial processes, which brings a new challenge-training heterogeneous models via multimodal data in federated learning.This paper proposes a novel federated learning approach called PyM-FL to tackle this challenge.The efficacy of PyM-FL is assessed through an experiment focused on the diagnosis of mild cognitive impairment.

Federated learning (FL) provides a promising solution for privacy-preserving intelligent applications by enabling collaborative model training with distributed clients without directly sharing their data.However, existing FL approaches often introduce a single point of failure, communication bottlenecks, and potential privacy risks at the central server.Additionally, real-world scenarios frequently involve data with multiple modalities.Although multimodality offers richer information for learning, current FL approaches lack the capability to fuse this information effectively.In this paper, we propose DeMoFed, a novel federated framework that facilitates multimodality fusion in a fully decentralized environment while aiming to reduce communication overhead.

Automatic monitoring and evaluation of chronic wounds usually requires massive labeled data sets for segmentation training.Because of the high cost of time and labor, these data are usually difficult to obtain.In order to improve the segmentation effect of the wound image with a small number of labeled samples for training, this paper proposes an image segmentation framework based on self-supervision, summarize a relatively optimal pre-training task for chronic wound image segmentation, minimizes the redundancy between the symmetric network projection output by learning the feature information of two views generated by the same chronic wound image under different distortion transformation, and finally learns valuable knowledge that is conducive to the downstream wound image segmentation task.In addition, this framework also optimize the network structure and loss of the segmentation model.The experimental results show that after self-supervised learning pretraining with a full amount of unlabeled data, the segmentation framework can achieve significant improvement in precision(up to 8%), recall(up to 5%), and MIoU(up to 9%) by fine-tuning with only a small amount of labeled data, This can provide a clear optimization direction for the application of self-supervised learning to specific image segmentation.

Alzheimer's disease (AD) is a neurodegenerative disease that severely affects cognition, memory, and behavior and is incurable.Mild cognitive impairment (MCI) is a clinical precursor to AD, and early diagnosis of AD is essential for the prevention and intervention of disease progression.The hippocampus is one of the first brain regions affected by AD, and therefore structural magnetic resonance images (sMRI) are commonly used to measure the shape and volume of the hippocampus.In this paper, we propose a federal deep forest model called FeDeFo for calculating hippocampal volume using sMRI images to achieve AD classification.Firstly, to effectively protect data privacy, we use a federated learning framework to collaboratively train a gradient boosting decision tree (GBDT) model based on the local data of each client.In addition, to address the data discrepancy between clients, we introduce a deep forest model to exploit the local data beyond local interactions further and fuse it with the federally trained GBDT to personalize the model for each client.The experiments demonstrate that our proposed approach is able to personalize the model while protecting the data privacy of each client, providing a new idea for AD classification.

Nowadays, chronic wounds have become an increasingly heavy healthcare burden.Therefore, wound classification is the most crucial task in wound diagnosis, which directly affects whether the treatment plan is optimal.This paper proposes a self-attention embedded residual network, or SARNet for short, which takes wound images as input and categorizes them into six types, i.e., burn wounds, surgical wounds, venous lower limb ulcers, pressure ulcers, diabetic foot ulcers, and normal skin.The classification accuracy of SARNet satisfactorily exceeds 80% mainly because its residual structure enhances the feature representation, and its built-in self-attention mechanism enables the global reference.

As an early stage of Alzheimer's disease (AD), the accurate detection of mild cognitive impairment (MCI) is very important for its early intervention. Due to the small between-group differences, it is easy to cause misclassification of MCI. Intrigued by the idea of metric learning, a novel framework based on cross-modal prototype generation is proposed to reduce the misclassification rate of MCI. First, the single modal prototypes are generated from magnetic resonance images (MRI), and then the auxiliary modal is used to correct the original prototypes and generate new cross-modal prototypes. An adaptive mechanism that can dynamically adjust the proportion of the auxiliary modal data is added in the proposed framework to improve the practicability of the cross-modal prototype framework. Experiments show that the cross-modal prototype generation framework can significantly reduce the misclassification rate of MCI by 40.3%.

At present, deep learning-based medical image segmentation algorithms have achieved fast and accurate semantic segmentation. However, their segmentation accuracy is still challenging to reach the clinical use standard, requiring further refinement by medical experts. Therefore, some researchers have turned their attention to interactive segmentation methods, which introduce human interaction to obtain information gain. Such methods model the dynamics of the image annotation process state and can effectively improve the segmentation accuracy under the interaction of medical experts. In this paper, we put forward a novel framework for the interactive segmentation of medical images based on deep reinforcement learning, namely FISMI-DRL. The experimental results demonstrate that our model achieves high segmentation accuracy and interaction efficiency.

Recent studies have shown that spontaneous speech can be exploited for cognitive decline screening in individuals with Alzheimer’s dementia. However, as the intermediate state between healthy control (HC) and Alzheimer’s disease (AD), mild cognitive impairment (MCI) is challenging to be distinguished from the other two. In order to tackle the problem, this paper proposes a two-stage metric learning approach. Each stage takes distinct acoustic features as input to a specific deep neural network. Moreover, we present an online triplet generator that can maximize sample utilization efficiency by investigating the decorrelation among samples. Finally, the experimental results prove that our proposed approach can significantly improve the accuracy of MCI detection and hence perform an excellent classification between subjects with AD, HC, and MCI.

The Mini Mental State Examination, referred to as MMSE, is a screening tool for cognitive dysfunction in the elderly, and it is also one of the most influential screening tools for cognitive impairment.It is usually managed by a well-trained doctor, but this is time-consuming and expensive.An effective method is to detect whether cognitive function has declined through the conversation between them.From the perspective of acoustics and linguistics, using 108 subjects provided by the Alzheimer's Dementia Recognition through Spontaneous Speech (ADReSS) 2020 Challenge, using speech to predict the MMSE score, the acoustic Root Mean Squared Error (RMSE) is 5.49.The RMSE in linguistics is 4.51.Integrating the acoustic model and the linguistic model, and assigning different weight ratios to their final predicted scores, the RMSE is 4.18.

The presence of different malicious regions on a single breast reveal some necessary information for breast cancer early detection. In current computer-aided diagnosis models, different lesions contained in a single mammogram are not detected and segmented individually. Therefore, the multidetection and segmentation of the breast lesions can help the radiologists for an accurate diagnosis. This study aims to develop a model based on regional learning technique and RoI-based Convolutional neural network (CNN), which is known as Masked Regional Convolutional Neural Network embedded with Feature Pyramid Network. By using Mask RCNN-FPN, we can handle multi-detection, instance segmentation, and classification simultaneously. FPN extracts semantic features at different resolution scales and it can exhibit lesions at multiple scales. The training and testing of the model are performed on the DDSM and Inbreast respectively. In comparison, this model achieved mean average precision 0.84 for multi-detection and segmentation and 91% overall accuracy performance over SegNet and U-Net CNN encoder and decoder segmentation architecture.

The multiple UUVs(unmanned underwater vehicles) can maintain a certain formation and complete the designated route cruise mission by using the path tracking controller. According to the problem of multiple UUV path tracking under discrete information conditions, a path tracking controller based on virtual leader is proposed in this paper considering communication delays. Firstly, the coupled nonlinear UUV kinematics and dynamics equations are transformed into second-order affine forms by using feedback linearization method, and the mathematical model under discrete time is given. Secondly, the path tracking controller is designed in the case of bounded time delays, and the sufficient conditions are analyzed and proposed based on the matrix theory. Finally, the effectiveness of the proposed algorithm is proved by the simulation experiments, all trajectories of each UUV converge to the desired trajectories under the specific formation.

One of the major steps for opinion mining is to extract product features. The vast majority of existing approaches focus on explicit feature identification, few attempts have been made to identify implicit features in reviews, however; people tend to express their opinions with simple structures and brachylogies, which lead to more implicit features in reviews. By analyzing the characteristics of product reviews in Chinese on the Internet, this paper proposes a novel context-based implicit feature extracting method. We extract the implicit features according to the opinion words and the similarity between the product features in the implicit features' context. We also build a matrix to show the relationship between opinion words and product features, then use a new algorithm to filter the noises in the matrix. Experiments show that our method provides higher accuracy in extracting the implicit features.

Trust-aware recommender system (TARS) can provide more relevant recommendation and more accurate rating predictions than the traditional recommender system by taking the trust network into consideration. However, most of the trust-aware collaborative filtering approaches do not consider the influence of contextual information on rating prediction. To the opposite, context-aware matrix factorization approaches as we know do not take trust information into consideration. In this paper, we propose two Trust-based Context-aware Matrix Factorization (TCMF) approaches to fully capture the influence of trust information and contextual information on ratings. We integrate both trust information and contextual information into the baseline predictors (user bias and item bias) and user-item-context-trust interaction. Evaluations based on a real dataset and three semi-synthetic datasets demonstrate that our approaches can improve the accuracy of the trust-aware collaborative filtering and the context-aware matrix factorization models by at least 10.2% in terms of MAE.

Context-aware recommender system (CARS) can provide more accurate rating predictions and more relevant recommendations by taking into account the contextual in-formation. Yet the state-of-the-art context-aware matrix factorization approaches only consider the influence of con-textual information on item bias. Tensor factorization based Multiverse Recommendation deals with the contextual in-formation by incorporating user-item-context interaction into recommendation model. However, all of these approaches cannot fully capture the influence of contextual information on the rating. In this paper, we propose two improved context-aware matrix factorization approaches to fully capture the influence of contextual information on the rating. Both of the baseline predictors (user bias and item bias) and user-item-context interaction are fully concerned. Experimental results on three semi-synthetic datasets and one real world dataset show that the two proposed approaches outperform Multiverse Recommendation and the state-of-the-art context-aware matrix factorization methods in prediction performance.