Uncertainty Quantification (UQ) is critical for detecting hallucinations in black-box Large Vision-Language Models (LVLMs). However, prevailing methods like Discrete Semantic Entropy (DSE) are unreliable, as their scores are primarily dominated by the number of semantic clusters. This renders them incapable of distinguishing between benign semantic ambiguity (varied but coherent responses) and severe belief conflict (contradictory responses). We address this limitation by proposing a novel framework rooted in Dempster-Shafer theory of evidence, built on the premise that not all inconsistency is equal. Our method decomposes uncertainty into two complementary metrics: Belief Divergence, which quantifies ambiguity by measuring the separation between viewpoints, and Belief Conflict, which captures direct logical contradictions. Extensive experiments demonstrate that our framework provides a more reliable measure of uncertainty.

Trusted multi-view classification (TMC) aims to improve prediction reliability by integrating evidence from multiple views. Existing TMC methods extract evidence from single view and use a regularization term to shape the evidence distribution. However, existing methods typically enforce a uniform regularization objective across all views, overlooking critical view-specific biases: intra-view class ambiguity caused by confusable features and inter-view quality disparities reflected in evidence uncertainty. To address these issues, we propose an adaptive regularization strategy that enhances robustness on two levels. At the intra-view level, it quantifies feature ambiguity to apply targeted relaxation to confusable classes, preventing over-penalization of inherent uncertainty. At the inter-view level, it evaluates relative view quality to impose stronger constraints on unreliable views and suppress noise from low-quality ones. Extensive experiments across multiple benchmarks demonstrate the superiority and reliability of the proposed method.

Active retrieval-augmented generation (RAG) triggers external knowledge retrieval during generation based on model-side uncertainty signals to support knowledge-intensive, multi-hop reasoning. However, existing methods often retrieve only after producing a complete answer, failing to surface and fill information gaps in time; moreover, relying on a single internal signal as the trigger cannot adequately capture the multifaceted nature of uncertainty. We therefore propose a conflict-aware active RAG framework. We first decompose complex questions into a sequence of step-level sub-problems. At each step, we quantify local distributional uncertainty via a sliding-window peak token entropy, and estimate cross-sample consensus via the variation ratio computed over multiple Monte Carlo samples. After calibrating both signals onto a probabilistic scale, we quantify their conflict using a symmetric, bounded divergence over Bernoulli parameters, and fuse the three quantities into a single uncertainty score that gates retrieval. Experiments demonstrate the effectiveness of our framework.

The fusion of multi-view medical images through deep neural networks is essential for boosting diagnostic precision in the field of medical image analysis. However, the reliability of these diagnostic results is often compromised by imperfections in image views, manifested as noise, artifacts, and data deficits arising from inconsistent diagnostic frequencies. These issues introduce a significant risk when merging medical views in a clinical setting. To address these problems, we introduce the Reliability-Enhanced Multi-view Network (REMNet), a novel framework designed to tackle two critical challenges: 1) reducing misclassification and uncertainty from imperfect view integration, and 2) improving the reliability and interpretability of multi-view medical image predictions. Specifically, REMNet merges information from multiple views into a coherent evidence framework and incorporates a Dirichlet prior within our predictive model to more accurately estimate confidence in predictions. Coupled with a robust fusion strategy and a precise confidence calibration process, REMNet consolidates the diverse strengths of various medical imaging views, reduces the impact of view imperfections, and enhances the reliability of medical imaging diagnostics. The superiority of REMNet is validated through comprehensive theoretical analysis and empirical experiments on multi-view medical image datasets across different modalities.

Open World Object Detection (OWOD) faces a fundamental dilemma: maintaining a stable representation for known classes while reserving flexible space for discovering unknown objects. Existing methods, while improving recall, often fail to assign discriminative confidence scores to unknown instances, resulting in critically low Average Precision (U-AP) and representation degradation during incremental learning. To remedy this, we propose the Evidential Prior Guided Neural Collapse (ENC) framework. ENC unifies representation learning and uncertainty quantification via a Geometric-Evidence Coupling mechanism. Unlike previous approaches, we map evidential support directly to the angular alignment with Simplex Equiangular Tight Frame (ETF) prototypes. Theoretically, the evidential prior functions as a geometric regularizer: it maximizes equiangular separation for confident known samples, while constraining ambiguous queries to approximate an isotropic uniform distribution via distributional regularization. Furthermore, to mitigate decision conflicts in self-supervised learning, we propose a dissonance-aware objectness optimization strategy that mines informative samples near the decision boundary. Extensive experiments on M-OWODB and S-OWODB benchmarks demonstrate that ENC sets a new state-of-the-art. Notably, it achieves a significant improvement in unknown class discovery, boosting U-AP from ≈ 1% to 9.2%, while exhibiting superior robustness against catastrophic forgetting in challenging incremental scenarios.

In the field of computer-aided diagnosis, particularly for tumor diseases, segmentation is a prerequisite and primary step. Multi-modality images become essential for achieving accurate segmentation, which offer critical insights beyond the limitations of single-modality data. However, different modalities and images may suffer from different types of data imperfection, such as intensity non-uniformity, motion artifact, and low quality due to hardware limitations, which challenge image segmentation algorithms. To address this challenge, we propose a Reliable Evidential Discounting Network (REDNet), which is composed of three main modules: 1) the Intra-modality Consistency Evaluation Module (ICEM) measuring the data cohesion within the same modality; 2) the Cross-modality Difference Aggregation Module (CDAM) identifing data discrepancy across modalities; 3) the Discounting Fusion Module (DFM) processing the multi-modality evidence by applying discounting strategies to fuse the data. This approach maintains segmentation accuracy by effectively integrating multi-modality evidence, while discounting the influence of lower-quality data, ensuring reliable results despite the presence of image imperfections. We evaluated REDNet on two distinct datasets, BRATS2021 and an in-house pancreas dataset from Changhai Hospital. REDNet outperforms other methods, particularly in scenarios with imperfect image sources, and achieves reliable results in multi-modality tumor segmentation.

Multi-view classification based on evidence theory aims to enhance result reliability by effectively quantifying prediction uncertainty at the evidence level, particularly when dealing with low-quality views. However, these methods face limitations in real-world applications due to the sensitivity of estimated uncertainty to view distribution, leading to two main issues: 1) difficulty in making clear judgments about whether to trust predictions based on vague uncertainty scores, and 2) the potential negative impact of integrating information from low-quality views on multi-view classification performance. Both limitations compromise the reliability of multi-view decisions. To address these challenges, we introduce an adaptive rejection mechanism based on estimated uncertainty, which is free of data distribution constraints. By integrating this adaptive rejection mechanism into the fusion of multiple views, our method not only indicates whether predictions should be adopted or rejected at the view level but also enhances classification performance by minimizing the impact of unreliable information. The effectiveness of our method is demonstrated through comprehensive theoretical analysis and empirical experiments on various multi-view datasets, establishing its superiority in enhancing the reliability of multi-view classification.

The automatic generation of medical imaging reports has significant research value, as it can alleviate the heavy workload of radiologists and reduce diagnostic biases. Existing methods mostly focus on text fluency, ensuring that generated reports align with ground truth. However, this leads to a neglect of the effective integration of clinical knowledge, resulting in poor clinical accuracy. To address this limitation, we propose a novel framework that incorporates clinical knowledge to improve clinical accuracy and report generation. Specifically, we introduce Multi-Expert Knowledge Enhanced Prompt Learning (KEP), which generates more reliable disease diagnostic prompts by integrating clinical knowledge, thereby enhancing clinical accuracy. Additionally, Knowledge-Enhanced Feature Learning (KEF) enhances image features with clinical knowledge, enabling the decoder to incorporate more comprehensive medical information. By integrating these knowledge-driven prompts and features into the generation model, our approach ensures clinical relevance and textual coherence. Experimental results demonstrate that our framework outperforms previous methods across multiple benchmarks, achieving superior clinical accuracy and linguistic fidelity.

Clinical disease classification naturally follows a hierarchical reasoning process, where clinicians first distinguish coarse diagnostic categories (e.g., benign vs. malignant) before identifying specific subtypes to enhance diagnostic accuracy. While multi-view imaging offers complementary perspectives, most existing machine learning approaches perform flat decision fusion, ignoring this intrinsic hierarchy and handling inter-view conflicts and uncertainty primarily at the decision-fusion level. To address this gap, we propose the Hierarchical Multi-view Evidence Aggregation Network (H-MEAN), an end-to-end architecture that explicitly mimics clinical decision-making process. H-MEAN consists of three key modules: (i) Multi-View Fusion, which jointly encodes heterogeneous imaging views to capture complementary features; (ii) Hierarchical Tree Alignment, which structurally propagates diagnostic cues across granularity levels to enforce clinical consistency; and (iii) Hierarchical Evidential Aggregation, which leverages evidential deep learning to quantify uncertainty throughout the diagnostic hierarchy. Extensive experiments on two multi-view medical datasets demonstrate that H-MEAN consistently outperforms competitive methods, achieving higher classification accuracy, particularly in challenging settings with missing or noisy views.

Medical images such as CT and PET play a significant role in cancer diagnosis, often accompanied by diagnostic reports that describe findings across multiple organs. These multimodal data sources offer clinicians detailed information to support comprehensive and accurate diagnoses. However, in the context of computer-aided diagnosis (CAD), the inherent noise and potential inconsistencies across different imaging modalities can introduce confounding information, thereby reducing diagnostic accuracy. To better extract disease-relevant features from full-body medical images and reports, we propose a causality-driven multi-modal alignment method. This approach extracts causal features from comprehensive medical images and structural reports, enabling machine learning models to focus on specific disease classification tasks. For both image and text modalities, we designed a causality-enhancement method to extract causal features from individual modalities respectively and align crossmodal features in fusion stage. Our approach provides a novel causal view on computer aided medical classification tasks.

Diffusion-Weighted Imaging (DWI) is essential in prostate cancer diagnosis, yet multi-b-value acquisition increases scan time and susceptibility to motion artifacts. Traditional interpolation methods fail to model complex tissue diffusion, and existing deep learning models often compromise lesion contrast and structural fidelity. We propose a novel framework that integrates latent diffusion models with a mask-guided lesionpreserving (MGLP) mechanism to synthesize low-b-value DWI from high-b-value inputs. The model leverages VQ-VAE-based discrete latent representations for feature alignment and employs lesion masks as spatial priors to preserve diagnostic regions during diffusion. A dynamic feature fusion strategy further enhances anatomical-functional consistency. Experiments on clinical datasets demonstrate superior performance over state-of-the-art methods in PSNR, SSIM, and especially lesion-specific SSIMROI metrics. Our approach effectively reduces the need for multi-b-value scans without sacrificing diagnostic reliability, showing potential as a supporting approach to reduce b-value scanning requirements in prostate MRI.

Multi-modal retinal image registration is critical for clinical diagnosis. Conventional methods often neglect the reliability of keypoint matching, which may lead to registration errors that compromise clinical trustworthiness. To address this challenge, we propose a novel trusted retinal image registration method based on Evidential Deep Learning (EDL) with uncertain keypoint matching. First, the detector networks are employed to measure the confidence and uncertainty of each keypoint within the image pairs. Based on the measure, the top k confident keypoints are selected from corresponding images respectively. Second, the descriptor networks are used to extract the features of the keypoints and compute the similarity matrix for keypoint pairs. Crucially, we utilize Dempster-Shafer theory to fuse the uncertainties of keypoints into the similarity calculation, which facilitates prioritizing high-confident keypoint correspondences while suppressing unreliable matches. Finally, based on the similarity matrix with uncertainty measure, a PROSAC-based registration algorithm is implemented to achieve robust keypoint alignment. Experiments on the FIRE and PRIME-FP20 datasets demonstrate that the proposed method achieves state-of-the-art performance. Ablation studies further validate the effectiveness of uncertainty fusion in image registration, yielding a significant improvement of 13.8% in Mean Landmark Error AUC{@}6.25 compared to the previous best method. Besides enhancing accuracy, through measuring the uncertain keypoint matching, the proposed method can produce trusted retinal image registration results for clinicians.

Cross-modal retrieval aims to match videos and texts by mapping them into a shared feature space. Most existing approaches achieve alignment through contrastive learning based on one-to-one supervised pairs. However, these methods rely too much on supervised signals and do not fully use the unsupervised semantic relationships within each modality. As a result, samples that are semantically similar may be spread in the shared space, which hurts retrieval performance. To solve this problem, we propose a method called Leveraging Intra-Modal Consistency for Cross-Modal Alignment and Retrieval (LICA). Our method introduces a consistency constraint between intra-modal similarities and cross-modal similarity distributions. In this way, samples that are close in meaning stay closer together in the shared space. Experiments on standard text-video retrieval benchmarks show that LICA helps optimize the distribution of the cross-modal feature space and improves retrieval accuracy.

Medical image diagnosis inherently involves uncertainty due to artifacts, occlusions, and ambiguous visual patterns, often leading to high inter-observer variability. While deep neural networks offer strong predictive performance, their outputs tend to be overconfident and poorly calibrated, limiting their clinical reliability. We propose Conformal Focal Loss (CFL), a principled approach that leverages the focal loss and the statistical validity of conformal prediction to better characterize diagnostic uncertainty. By emphasizing hard or ambiguous examples, CFL enables more accurate estimation of both predictive confidence and ambiguity. We evaluate CFL on diagnostic tasks using both clean and noise-augmented datasets, demonstrating its ability to effectively identify uncertain cases while maintaining robust classification performance under label noise.

Trusted multi-view classification (TMVC) addresses variations in data quality by evaluating the reliability of each view based on prediction uncertainty at the evidence level, reducing the impact of low-quality views commonly encountered in real-world scenarios. However, existing TMVC methods often struggle to maintain robustness during testing, particularly when integrating noisy or corrupted views. This limitation arises because the evidence collected by TMVC may be unreliable, frequently providing incorrect information due to complex view distributions and optimization challenges, ultimately leading to classification performance degradation. To enhance the robustness of TMVC methods in real-world conditions, we propose a generalized evidence filtering mechanism that is compatible with various fusion strategies commonly used in TMVC, including Belief Constraint Fusion, Aleatory Cumulative Belief Fusion, and Averaging Belief Fusion. Specifically, we frame the identification of unreliable evidence as a multiple testing problem and introduce p-values to control the risk of false identification. By selectively down-weighting unreliable evidence during testing, our mechanism ensures robust fusion and mitigates performance degradation. Both theoretical guarantees and empirical results demonstrate significant improvements in the classification performance of TMVC methods, supporting their reliable application in challenging, real-world environments.

Trustworthy multi-view classification is essential for safety-critical applications, yet existing methods often fail when view quality degrades at test time due to noise or domain shifts. We propose a novel test-time adaptive fusion framework that leverages Conformal Prediction (CP) to transform pre-trained single-view classifier outputs into robust evidence representations. By integrating CP with Dempster-Shafer Theory, our approach dynamically generates well-calibrated evidence for each view, enabling reliable fusion under arbitrary view conditions without retraining. Extensive experiments demonstrate that our method achieves competitive performance on clean data and significantly outperforms prior methods in the presence of view corruption, offering a practical solution for robust multi-view classification in real-world scenarios.

Gradient-based methods provide a principled framework for Active Learning (AL), effectively quantifying a sample’s informativeness through the uncertainty captured from the model’s internal dynamics. However, their effectiveness is often hindered by the failure to disentangle epistemic and aleatoric uncertainty. This can lead to the suboptimal selection of noisy outliers over truly informative samples. To address this, we propose Dual-View Gradient Probes (DVGradProb), a framework that disentangles these uncertainties by probing gradients from two complementary spaces. Specifically, epistemic uncertainty is captured from the parameter space and introduce a novel metric for aleatoric uncertainty from the feature space. These metrics are integrated into a ratio-based acquisition function designed to filter out noisy outliers while prioritizing truly informative samples, those with high epistemic but low aleatoric uncertainty. Experiments on MNIST, CIFAR-10, and SVHN demonstrate that DVGradProb robustly outperforms baselines, validating the effectiveness of this uncertainty disentanglement strategy.

Uncertainty estimation is crucial for ensuring the reliability of machine learning models in safety-critical applications. Evidential Deep Learning (EDL) offers a principled framework by modeling predictive uncertainty through Dirichlet distributions over class probabilities. However, existing EDL methods predominantly rely on level-0 hard labels, which supervised a uncertainty-aware model with full certainty. We argue that hard labels not only fail to capture epistemic uncertainty but also obscure the aleatoric uncertainty arising from inherent data noise and label ambiguity. As a result, EDL models often produce degenerate Dirichlet distributions that collapse to near-deterministic outputs. To overcome these limitations, we propose a vicinal risk minimization paradigm for EDL by incorporating level-1 supervision in the form of vicinally smoothed conditional label distributions. This richer supervision exposes the model to local label uncertainty, enhancing aleatoric uncertainty quantification, while also mitigating the degeneration of the Dirichlet distribution into a Dirac delta function, thereby improving epistemic uncertainty modeling. Extensive experiments show that our approach consistently outperforms standard EDL baselines across synthetic datasets, covariate-shifted out-of-distribution generalization tasks, and out-of-distribution detection benchmarks, providing more reliable uncertainty estimates.

Effective exploration remains a core challenge in deep reinforcement learning, particularly in sparse-reward environments where agents receive limited feedback and traditional value-based methods often suffer from overestimation bias. Standard approaches, such as Deep Q-Networks, generate point estimates of Q-values and cannot represent the epistemic uncertainty introduced by limited and sparse interactions, which can lead to inefficient exploration and suboptimal policy learning. To address these limitations, we propose an uncertainty-aware exploration framework based on Dempster-Shafer theory to improve policy learning in sparse-reward scenarios. We model the uncertainty of the Q-value by constructing belief-plausibility intervals. These intervals are derived through basic probability assignments over action subsets, which models the confidence of value estimates. A credibility-driven decision mechanism is introduced to incorporate these intervals into action selection. Furthermore, we design a fusion strategy that combines the uncertainty bounds with traditional Q-values to form a unified hybrid value representation, which is used for both policy updates and target computation. Experimental results on multiple discrete control tasks with sparse reward demonstrate that our method significantly improves both the convergence speed and the exploration efficiency compared to standard baselines.

In recent years, multi-view learning has been widely applied in medical imaging. However some low-quality data obtained may contain noise and artifacts, leading to unreliable decisions. Several trustworthy multi-view learning methods have been proposed to assess the uncertainty of decisions. In the diagnosis of pancreatic tumor subtypes (PTS), the subtle differences between subtypes make them difficult to distinguish. Without the guidance of clinical prior knowledge, these method often struggle to capture the key features specific to each subtype. To address these challenges, we propose a novel trustworthy multi-view learning method. By analyzing the potential malignancy risks of PTS, we classify all subtypes into three broad categories. This coarse-grained vague classification informed by clinical prior knowledge is introduced into the neural network as auxiliary information to generate augmented opinions. Additionally, to compensate for the lack of global features, we introduce a global view to supplement the feature space and strengthen multi-view learning. Finally, using an uncertainty-based combination rule, the augmented opinions extracted from multiple views are reliably combined for the final diagnosis. We conduct a series of experiments comparing our approach with the state-of-the-art multi-view learning methods. The experimental results validate the effectiveness of our proposed method, demonstrating its superior performance and trustworthiness.

Accurate segmentation of brain tumors is crucial for effective diagnosis and treatment planning. Magnetic Resonance Imaging (MRI) with multiple modalities provides comprehensive biological information for tumor diagnosis. However, the absence of certain modalities degrades segmentation performance. To address this challenge, we propose an Evidence Harmonization Network (EHNet), which is designed for brain tumor segmentation in scenarios with missing modalities. Firstly, we use subjective logic to model uncertainty, replacing missing modality evidence with averaged evidence from available modality. Secondly, we introduce the Evidence Harmonization Module, which captures the distinction and consensus among modalities and calculates a harmonization factor. It enables the network to dynamically extract effective features in missing modalities scenarios and fully utilizes the strengths of each available modalities for final segmentation. Finally, the Dynamic Fusion Module adjusts the evidence with the harmonization factor and fuses it to produce the final segmentation results. We validate our method on the widely used Brats2018 dataset, demonstrating its superior performance compared to existing methods in missing modalities scenarios. Our approach not only achieves accurate segmentation results but also provides reliable uncertainty estimation.

Given a specific class space, open set recognition (OSR) requires a model to not only accurately classify data samples belonging to that class space, but also to identify samples from unseen classes. Based on prototype learning, existing OSR methods seek to build a separable feature space in which the unseen classes are as far away as possible from the given classes. However, these methods may overly focus on either the unseen classes or the given classes, resulting in unstable performance and lack of robustness. To handle this challenge, a robust OSR learning framework, named dual prototype learning is proposed. During the training phase, a set of dual prototypes is created for each given class, where one prototype (intra-class prototype) is used to form a compact intra-class feature representation, and the other prototype (anti-class prototype) is used to form a space that does not belong to the class by pushing the anti-class prototype away from the intra-class prototype. In the inference phase, our method can achieve a more stable and robust decision-making process by simultaneously measuring the distance between the sample and the intra-class prototypes, as well as the distance between the sample and the anti-class prototypes. Qualitative quantitative experiments demonstrate that the proposed method outperforms existing methods in terms of classification performance and robustness.

Uncertainty-aware multi-view deep classification methods have markedly improved the reliability of results amidst the challenges posed by noisy multi-view data, primarily by quantifying the uncertainty of predictions. Despite their efficacy, these methods encounter limitations in real-world applications: 1) They are limited to providing a single class prediction per instance, which can lead to inaccuracies when dealing with samples that are difficult to classify due to inconsistencies across multiple views. 2) While these methods offer a quantification of prediction uncertainty, the magnitude of such uncertainty often varies with different datasets, leading to confusion among decision-makers due to the lack of a standardized measure for uncertainty intensity. To address these issues, we introduce Conformalized Multi-view Deep Classification (CMDC), a novel method that generates set-valued rather than single-valued predictions and integrates uncertain predictions as an explicit class category. Through end-to-end training, CMDC minimizes the size of prediction sets while guaranteeing that the set-valued predictions contain the true label with a user-defined probability, building trust in decision-making. The superiority of CMDC is validated through comprehensive theoretical analysis and empirical experiments on various multi-view datasets.

Evidential Deep Learning (EDL), grounded in Evidence Theory and Subjective Logic (SL), provides a robust framework to estimate uncertainty for out-of-distribution (OOD) detection alongside traditional classification probabilities.However, the EDL framework is constrained by its focus on evidence that supports only single categories, neglecting the other collective evidences that could corroborate multiple in-distribution categories. This limitation leads to a diminished estimation of uncertainty and a subsequent decline in OOD detection performance.Additionally, EDL encounters the vanishing gradient problem within its fully-connected layers, further degrading classification accuracy.To address these issues, we introduce hyper-domain and propose Hyper-opinion Evidential Deep Learning (HEDL). HEDL extends the evidence modeling paradigm by explicitly integrating sharp evidence, which supports a singular category, with vague evidence that accommodates multiple potential categories.Additionally, we propose a novel opinion projection mechanism that translates hyper-opinion into multinomial-opinion, which is then optimized within the EDL framework to ensure precise classification and refined uncertainty estimation.HEDL integrates evidences across various categories to yield a holistic evidentiary foundation for achieving superior OOD detection. Furthermore, our proposed opinion projection method effectively mitigates the vanishing gradient issue, ensuring classification accuracy without additional model complexity. Extensive experiments over many datasets demonstrate our proposed method outperforms existing OOD detection methods.

Accurately segmenting tubular structures, such as blood vessels or nerves, holds significant clinical implications across various medical applications. However, existing methods often exhibit limitations in achieving satisfactory topological performance, particularly in terms of preserving connectivity. To address this challenge, we propose a novel deep-learning approach, termed Deep Closing, inspired by the well-established classic closing operation. Deep Closing first leverages an AutoEncoder trained in the Masked Image Modeling (MIM) paradigm, enhanced with digital topology knowledge, to effectively learn the inherent shape prior of tubular structures and indicate potential disconnected regions. Subsequently, a Simple Components Erosion module is employed to generate topology-focused outcomes, which refines the preceding segmentation results, ensuring all the generated regions are topologically significant. To evaluate the efficacy of Deep Closing, we conduct comprehensive experiments on 4 datasets: DRIVE, CHASE_DB1, DCA1, and CREMI. The results demonstrate that our approach yields considerable improvements in topological performance compared with existing methods. Furthermore, Deep Closing exhibits the ability to generalize and transfer knowledge from external datasets, showcasing its robustness and adaptability. The code for this paper has been available at: https://github.com/5k5000/DeepClosing.

Graph-based clustering approaches, especially the family of spectral clustering, have been widely used in machine learning areas. The alternatives usually engage a similarity matrix that is constructed in advance or learned from a probabilistic perspective. However, unreasonable similarity matrix construction inevitably leads to performance degradation, and the sum-to-one probability constraints may make the approaches sensitive to noisy scenarios. To address these issues, the notion of typicality-aware adaptive similarity matrix learning is presented in this study. The typicality (possibility) rather than the probability of each sample being a neighbor of other samples is measured and adaptively learned. By introducing a robust balance term, the similarity between any pairs of samples is only related to the distance between them, yet it is not affected by other samples. Therefore, the impact caused by the noisy data or outliers can be alleviated, and meanwhile, the neighborhood structures can be well captured according to the joint distance between samples and their spectral embeddings. Moreover, the generated similarity matrix has block diagonal properties that are beneficial to correct clustering. Interestingly, the results optimized by the typicality-aware adaptive similarity matrix learning share the common essence with the Gaussian kernel function, and the latter can be directly derived from the former. Extensive experiments on synthetic and well-known benchmark datasets demonstrate the superiority of the proposed idea when comparing with some state-of-the-art methods.

Multi-view deep classification expects to obtain better classification performance than using a single view. However, due to the uncertainty and inconsistency of data sources, adding data views does not necessarily lead to the performance improvements in multi-view classification. How to avoid worsening classification performance when adding views is crucial for multi-view deep learning but rarely studied. To tackle this limitation, in this paper, we reformulate the multi-view classification problem from the perspective of safe learning and thereby propose a Safe Multi-view Deep Classification (SMDC) method, which can guarantee that the classification performance does not deteriorate when fusing multiple views. In the SMDC method, we dynamically integrate multiple views and estimate the inherent uncertainties among multiple views with different root causes based on evidence theory. Through minimizing the uncertainties, SMDC promotes the evidences from data views for correct classification, and in the meantime excludes the incorrect evidences to produce the safe multi-view classification results. Furthermore, we theoretically prove that in the safe multi-view classification, adding data views will certainly not increase the empirical risk of classification. The experiments on various kinds of multi-view datasets validate that the proposed SMDC method can achieve precise and safe classification results.

Fine-Grained Image Classification (FGIC) aims to classify images into specific subordinate classes of a superclass. Due to insufficient training data and confusing data samples, FGIC may produce uncertain classification results that are untrusted for data applications. In fact, FGIC can be viewed as a hierarchical classification process and the multilayer information facilitates to reduce uncertainty and improve the reliability of FGIC. In this paper, we adopt the evidence theory to measure uncertainty and confidence in hierarchical classification process and propose a trusted FGIC method through fusing multilayer classification evidence. Comparing with the traditional approaches, the trusted FGIC method not only generates accurate classification results but also reduces the uncertainty of fine-grained classification. Specifically, we construct an evidence extractor at each classification layer to extract multilayer (multi-grained) evidence for image classification. To fuse the extracted multi-grained evidence from coarse to fine, we formulate evidence fusion with the Dirichlet hyper probability distribution and thereby hierarchically decompose the evidence of coarse-grained classes into fine-grained classes to enhance the classification performances. The ablation experiments validate that the hierarchical evidence fusion can improve the precision and also reduce the uncertainty of fine-grained classification. The comparison with state-of-the-art FGIC methods shows that our proposed method achieves competitive performances.

Real-world classification tasks often show an extremely imbalanced problem. The extreme imbalance will cause a strong bias that the decision boundary of the classifier is completely dominated by the categories with abundant samples, which are also called the head categories. Current methods have alleviated the imbalanced impact from mainly three aspects: class re-balance, decoupling and domain adaptation. However, the existing criterion with the winner-take-all strategy still leads to the crowding problem in the eigenspace. The head categories with many samples can extract features more accurately, but occupy most of the eigenspace. The tail categories sharing the rest of the narrow eigenspace are too crowded together to accurately extract features. Above these issues, we propose a novel T-distributed spherical metric for equalized eigenspace in the imbalanced classification, which has the following innovations: 1) We design the T-distributed spherical metric, which has the characteristics of high kurtosis. Instead of the winner-take-all strategy, the T-distributed spherical metric produces a high logit only when the extracted feature is close enough to the category center, without a strong bias against other categories. 2) The T-distributed spherical metric is integrated into the classifier, which is able to equalize the eigenspace for alleviating the crowding issue in the imbalanced problem. The equalized eigenspace by the T-distributed spherical classifier is capable of improving the accuracy of the tail categories while maintaining the accuracy of the head, which significantly promotes the intraclass compactness and interclass separability of features. Extensive experiments on large-scale imbalanced datasets verify our method, which shows superior results in the long-tailed CIFAR-100/-10 with the imbalanced ratio IR = 100/50. Our method also achieves excellent results on the large-scale ImageNet-LT dataset and the iNaturalist dataset with various backbones. In addition, we provide a case study of the real clinical classification of pancreatic tumor subtypes with 6 categories. Among them, the largest number of PDAC accounts for 315 cases, and the least CP has only 8 cases. After 4-fold cross-validation, we achieved a top-1 accuracy of 69.04%.

Fine-Grained Medical Image Classification (FGMIC) aims to identify disease subclasses within the corresponding metaclass. Due to insufficient labeled images and confusing image samples, the accuracy of FGMIC is limited and the trustworthiness of the model is also affected. In this paper, we utilize evidence theory to measure prediction uncertainty and improve the trustworthiness of FGMIC through multiple evidence fusion. Specifically, we consider FGMIC as a hierarchical classification process. At each layer, we construct an evidential classifier to extract classification evidence. Evidence extracted from all layers forms multi-grained evidence. Then, multi-grained evidence are fused through the Dirichlet hyper-PDF, so that evidence of coarse-grained layer classes can be used to enhance the corresponding evidence of fine-grained layer classes. Moreover, the scanned 3D medical image of a patient can generally be divided into three 2D views, with different views containing different features and uncertainties of the pathological region. Inspired by this, the evidential classifier of each layer is split into three sub-evidential classifiers, where one sub-evidential classifier is built on a view. Then, classification evidence from different views is fused using uncertainty-weighted fusion. Experiments on two cancer subtype classification tasks validate that multiple evidence fusion can not only improve prediction accuracy, but also reduce uncertainty and improve the trustworthiness.

Object Detection (OD) can be viewed as a multi-objective task to achieve object localization and class recognition. With the rapid development of the deep neural networks (DNNs), on the one hand, the performance of OD has been significantly improved by relying on the high-quality feature extraction and representation of DNNs. On the other hand, it can be challenging to accurately detect and recognize objects with non-salient or confusing features. In this paper, we propose an efficient and pluggable OD method by using attention mechanism to solve these issues from two aspects. Firstly, we exploit the semantic relationship between objects as a prior knowledge to reduce the incorrect recognition of objects with confusing features, where the relationship is encoded as an attention map by using a graph convolutional network, and then this attention map is used to reweight the feature intensities of objects belonging to different classes. Then, based on the feature map extracted from DNNs, we extract a sub-feature map containing foreground information and use this map to generate foreground attention map to improve the feature saliency of the objects. The qualitative and quantitative experimental results well verify the effectiveness of our method.

In recent years, significant progress has been made in 3D object detection. The focus of research has primarily been on improving the detection accuracy of models, however, neglecting their efficiency during actual deployment. Aiming at this issue, in this paper, we propose a multi-dimensional pruning method from the perspectives of data and model. Specifically, given the input data represented by the voxel grid, we first measure the voxel importance and propose an importance-based sampling module to sparsify voxels while preserving informative ones. The model pruning is wrapped in the framework of weighted voxel distillation, where the student model is obtained by pruning the channels of teacher model and only the informative voxels in the teacher model are involved and transferred to the students. In addition, the proposed method can be seamlessly integrated into current voxel-based 3D detectors without any additional costs. Experimental results on the KITTI and ONCE datasets show that our method can achieve a reduction of over 80% in GFLOPs while maintaining superior performance.

The performance of k-means clustering is often degenerate when dealing with high-dimensional and noisy scenarios. In this study, an end-to-end robust clustering method with low-rank linear embedding techniques (RCLR) is presented in conjunction with k-means. Sparse coefficients and a space projection matrix can be simultaneously learned. The global structures and local neighborhood properties are well captured in the learning procedures. Both the processes of clustering and dimensionality reduction are realized at the same time. The notions of clustering, dimensionality reduction, low-rank representation, and local property preservation are seamlessly integrated into a unified model. The limitation of error accumulation encountered in the previous two-stage clustering framework involving low-rank representation can be alleviated. This is the first attempt to introduce both the global and local geometrical structures into k-means directly, as well L2,1-norm is used as a basic metric instead of the conventional F-norm to further improve the robustness and interpretation of the model. The superiority of the proposed RCLR method is demonstrated by extensive experiments completed on various well-known benchmark datasets.

Multi-view deep learning is performed based on the deep fusion of data from multiple sources, i.e. data with multiple views. However, due to the property differences and inconsistency of data sources, the deep learning results based on the fusion of multi-view data may be uncertain and unreliable. It is required to reduce the uncertainty in data fusion and implement the trusted multi-view deep learning. Aiming at the problem, we revisit the multi-view learning from the perspective of opinion aggregation and thereby devise a trusted multi-view deep learning method. Within this method, we adopt evidence theory to formulate the uncertainty of opinions as learning results from different data sources and measure the uncertainty of opinion aggregation as multi-view learning results through evidence accumulation. We prove that accumulating the evidences from multiple data views will decrease the uncertainty in multi-view deep learning and facilitate to achieve the trusted learning results. Experiments on various kinds of multi-view datasets verify the reliability and robustness of the proposed multi-view deep learning method.

Pre-training deep learning models on ImageNet and transferring the models to medical image applications facilitate to improve the medical image analysis and reduce the need for labeled medical image data. However, some images from ImageNet may be fundamentally different from medical images in feature representation and lead to the negative transfer effects. To deal with this issue, we propose a novel strategy based on evidence theory to select reliable instances from ImageNet for medical image domain adaptation. Specifically, we formulate an evidential mass function to measure the ignorance and reliability of the images from ImageNet with respect to the classification tasks of medical images. Through selecting reliable instances with low ignorance degree from ImageNet, we can enhance the transfer performances of deep neural networks in medical image domain adaptation. Moreover, the proposed data selection strategy is independent of specific learning algorithm and can be viewed as a common preprocessing technique. Numerical experiments on tomography images, X-Ray images, and ultrasound images are given to comprehensively demonstrate the effectiveness of the selection strategy.

As an unsupervised ensemble learning strategy, clustering ensemble combines multiple base clusterings into a high-quality one and has achieved successful applications in image analysis and data mining. However, extant clustering ensemble methods are ineffective to handle the data uncertainty in clustering consensus process, which may mislead to poor clustering ensemble results. To tackle the problem, we propose a stable clustering ensemble (SCE) method based on evidence theory (Dempster–Shafer theory) in this paper. Specifically, we construct a belief function of cluster membership to measure the uncertainty and stability of data instances in clustering ensemble and thereby implement the stable clustering ensemble algorithm. We test the proposed stable clustering ensemble method in the tasks of structural data clustering and image segmentation. The experimental results validate the proposed method is effective to process the uncertain data and produce high-quality data clusterings.

Labeling cerebral vessels requires domain knowledge in neurology and could be extremely laborious, and there is a scarcity of public annotated cerebrovascular datasets. Traditional machine learning or statistical models could yield decent results on thick vessels with high contrast while having poor performance on those regions of low contrast. In our work, we employ a statistic model as noisy labels and propose a Transformer-based architecture which utilizes Hessian shape prior as soft supervision. It enhances the learning ability of the network to tubular structures, so that the model can make more accurate predictions on refined cerebrovascular segmentation. Furthermore, to combat the overfitting towards noisy labels as model training, we introduce an effective label extension strategy that only calls for a few manual strokes on one sample. These supplementary labels are not used for supervision but only as an indicator to tell where the model keeps the most generalization capability, so as to further guide the model selection in validation. Our experiments are carried out on a public TOF-MRA dataset from MIDAS data platform, and the results demonstrate that our method shows superior performance on cerebrovascular segmentation which achieves Dice of 0.831±0.040 in the dataset.

The cross-modality generation is usually a challenging task, especially for the medical image with a huge distribution difference. Micro Computed Tomography (Micro-CT) has high resolution and sharp contrast, which is regarded as the gold standard of tooth structure in dentistry. But Micro-CT has to scan a single isolated tooth in vitro, it cannot provide any information before surgery. Cone Beam Computed Tomography (CBCT) is commonly used in clinical treatment, but its low resolution and blurred border often lead to the deficiency of accurate tooth structure for surgery, such as the pulp and crown. The modality conversion from CBCT to Micro-CT is very difficult due to the huge difference between two modality distributions. Most GAN-based models cannot cope with such distribution differences between modalities. Therefore, we propose a novel cross-modality method for generating Micro-CT from CBCT. First, we propose a variation-based distance to reflect the difference between the two modality distributions. Second, integrating the variation-based distance, we design an end-to-end neural network named Variational Synthesis network (VSnet) to generate the Micro-CT from CBCT. Finally, we compare our method with plentiful state-of-the-art image generation methods. The results show that our proposed method is superior to other existing methods on the Micro-CT generation task, with image quality index SSIM of 0.912 $(\pm 0.061)$, PSNR of 22.451 $(\pm 4.810)$, and NRMSE of 0.261 $(\pm 0.142)$. In terms of the generation of the tooth structure, we also achieve superior results in clinic application. For the Micro-CT generated by VSnet, it is easy to achieve precise segmentations of tooth pulp and crown through simple thresholding, which facilitates the diagnosis in dental clinic.

Bladder cancer staging is crucial for operation planning and cancer assessment. Deep Convolutional Neural Networks (DCNNs) have been widely used to classify the bladder tumor images to identify cancer stages. However, the pure image-based deep learning methods over depend on the labeled data training and neglect the clinical priors. Human doctors judge the stage of a bladder tumor through checking whether the tumor infiltrating into bladder wall. The clinical priors of tumor infiltration are helpful to improve the DCNN-based bladder cancer staging and make the predictions coincide with the law of medicine. To involve clinical priors into deep learning for cancer staging, we propose a DCNN model with prior evidence to classify medical images of bladder tumors. Specifically, we measure the degree of tumor infiltrating into bladder wall to construct the prior evidence and integrate the prior evidence into the image-based prediction with evidential deep neural networks. We analyze the learning objective and prove that the prior evidences consistent with the ground truth will certainly reduce the prediction error and variance produced by image-based neural networks. The experiments on bladder cancer MR images datasets validate that involving prior evidences is effective to improve the DCNN-based cancer staging.

Segmenting bladder tumors from MR images is important for the early detection and auxiliary diagnosis of bladder cancer. In recent years, Deep Convolutional Neural Networks (DCNNs) have been widely used for bladder tumor segmentation. However, the DCNN-based tumor segmentation methods generally over-depend on the model training with labeled data, and neglect the priors that can be learned from data or domain experts. The priors of tumor size and location are very helpful to improve the segmentation precision of bladder tumors, especially for the situation when the labeled images are insufficient to cover the diversity. To tackle the problem, we proposed a method to integrate priors into DCNN-based bladder tumor segmentation in this paper. The priors consist of both specific prior from data and general prior from domain knowledge. Specifically, we add a decoding path into the segmentation network to generate the specific prior, and formulate the general prior of domain knowledge to guide the generation of specific prior. Comprehensive experiments on massive bladder MR images clearly validate the effectiveness of the proposed bladder tumor segmentation method with priors.

Bladder cancer is a malignant disease with substantial morbidity and mortality. Bladder cancer staging is crucial to determine the effective treatments of bladder tumors in clinic. As to the superiority of feature learning, Deep Convolutional Neural Networks (DCNN) are widely used to predict the cancer stage based on medical images. However, most existing DCNN-based cancer staging methods are data-driven and neglect the domain knowledge and experiences of clinicians. Besides, the deep neural networks are short of model interpretability and may lead to risky diagnosis. To tackle the problems, we construct the diagnosis rules of bladder cancer staging based on the clinical experiences of tumor penetration into bladder wall. The diagnosis rules are extracted from Magnetic Resonance (MR) images and further integrated into DCNN for joint identification of tumor stage. The experiments validate that the integrated rules improve the model interpretability and guide DCNN to focus on the regions of tumor penetration and thereby produce precise prediction of cancer staging.

Neighborhoods form a set-level approximation of data distribution for learning tasks. Due to the advantages of data generalization and nonparametric property, neighborhood models have been widely used for data classification. However, the existing neighborhood-based classification methods rigidly assign a certain class label to each data instance and lack the strategies to handle the uncertain instances. The far-fetched certain classification of uncertain instances may suffer serious risks. To tackle this problem, in this article, we propose a novel shadowed set to construct shadowed neighborhoods for uncertain data classification. For the fuzzy-rough transformation in the proposed shadowed set, a step function is utilized to map fuzzy neighborhood memberships to the set of triple typical values {0, 1, 0.5} and thereby partition a neighborhood into certain regions and uncertain boundary (neighborhood shadow). The threshold parameter in the step function for constructing shadowed neighborhoods is optimized through minimizing the membership loss in the mapping of shadowed sets. Based on the constructed shadowed neighborhoods, we implement a three-way classification algorithm to distinguish data instances into certain classes and uncertain case. Experiments validate the proposed three-way classification method with shadowed neighborhoods is effective in handling uncertain data and reducing the classification risk.

Shadowed sets provide a meaningful description of information granules by abstracting the corresponding fuzzy sets into three categories: full acceptance, full rejection, and uncertain (represented by shadows). One of the main motivating points to derive shadowed sets from fuzzy sets is the determination and explanation of the separation thresholds based on a specific optimization mechanism. The available optimization objective functions are mainly discussed on semantic interpretations and their mathematical properties; constructive algorithms for optimal solutions have rarely been reported. In this paper, the continuous and convex properties of Pedrycz's optimization objective function to construct shadowed sets, as well as the existence and uniqueness of solution points, are analyzed in detail. It is demonstrated that different approximation region partitions would be generated even under the same optimization model, which requires further criteria to make the constructed shadowed sets well-defined. To address this limitation, the notions of passive and active constrained shadowed sets are introduced. A fast algorithm to obtain the proposed constrained shadowed sets is also designed based on the analyzed mathematical properties. Its performance is then illustrated by some typical fuzzy sets and some real data from the UCI repository.

The ensemble of feature selections facilitates to improve data generalization for learning tasks. However, existing feature selection ensemble methods have the following drawbacks. First, focusing on the numerical data, the works on the feature selection ensemble for symbolic and mixed-type data are very limited. Second, the voting-based ensemble strategies tend to select the top significant features but the consistency between the ensemble result and the diverse feature selections cannot be guaranteed. Aiming to handle these problems, we propose an feature selection ensemble method based on Analytic Hierarchy Process (AHP) in this paper. The AHP-based ensemble method can integrate diverse feature selections into a consistent one under the multiple criteria of feature discernibility and independence. Moreover, the ensemble methodology is helpful to implement the feature selection on distributed data and involve domain knowledge through extending criteria. Experimental results validate that the proposed ensemble method of feature selections is effective for symbolic data classification.

Most image segmentation methods based on active contour model are sensitive to the contour initialization. For the images of complex contents, it is difficult to initialize the contour properly and the biased initialization may lead to low-quality segmentation results. Aiming to tackle this problem, we propose an ensemble strategy to improve the contour-based segmentation. The optimal segmentation ensemble is obtained through maximizing the weighted mutual information between the probability distributions of multiple segmentation results. Experimental results validate that the ensemble of contour-based segmentation is robust to the biased initialization and produces stable and precise results for the images of complex contents.

Neighborhood Covering Reduction (NCR) is an effective tool to learn rules from structural data for classification. However, the existing neighborhood covering model is not robust enough. A neighborhood is constructed according to the nearest heterogeneous samples. This strategy over focuses on the boundary samples and makes the model sensitive to noise. To tackle this problem, we proposed a Rough Neighborhood Covering Reduction method (RCR) for robust classification. In RCR method, we construct an approximation of neighborhood based on rough sets and further design a reduction algorithm to filter out neighborhoods to form flexible covering of data space for classification. Abundant experiments verify the robustness of the proposed method, which achieves precise and stable classification results on noisy data.

With the information overload in web services, micro-blog has been increasingly providing as a media for end-users to express their opinions. The notable feature of micro-blog articles is prone to be a burst of popularity during a short period. In addition, diverse interests make users bored in redundant items in most recommender systems. Therefore, providing users with diverse popular micro-blogs that suit their interesting topics is an important issue. In this paper, depending on forwarding number and comment number of micro-blogs, an effective model for popularity prediction is proposed to discover popular topics. Then, a MaxMin diversity algorithm based on content distance and popularity density is proposed to discover top k micro-blogs. Finally, we design a diverse personalized popularity attention (DPPA) recommendation approach for target user. We conduct extensive experiments on large scale micro-blog datasets. The experimental results show that our proposed approach can satisfy user's requirements with a higher recall than personal attention methods.