Segmentation models based on deep neural networks demonstrate strong generalization for medical image segmentation. However, they often exhibit overconfidence or underconfidence, leading to unreliable confidence scores for segmentation masks, especially in ambiguous regions. This limits the reliability of fully automatic segmentation in ambiguous regions. Motivated by the learning-to-defer (L2D) paradigm, we introduce Deferred-Seg, a pixel-wise Human-AI collaborative deferral framework that learns whether to retain model predictions or defer difficult pixels or regions to expert branches. DeferredSeg extends the base segmentor with an aggregated deferral predictor and additional routing channels that dynamically route each pixel to either the base segmentor or human expert. To train this routing efficiently, we introduce a pixel-wise surrogate collaboration loss that supervises deferral decisions. In addition, we introduce a spatial-coherence loss that aligns the aggregated expert-assignment probability with the learned routing decisions, encouraging locally consistent deferral patterns. Beyond single-expert deferral, we further extend the framework to a multi-expert setting by introducing multiple experts with different reliability profiles for collaborative decision-making. To mitigate overloading or underutilization of individual experts, we further design a load-balancing penalty that discourages routing collapse and promotes more balanced utilization across expert branches. We evaluate DeferredSeg on multiple medical segmentation benchmarks using MedSAM and CENet as the base segmentors, and further compare it with an expert-assisted patch-correction baseline. Experimental results show that DeferredSeg consistently improves System-level segmentation performance while selectively deferring difficult pixels or regions to more reliable expert branches. Moreover, the proposed framework is backbone-compatible and can be instantiated with different segmentation architectures that provide intermediate feature representations.
Given a graph F, a hypergraph is called a Berge-F if it can be obtained by expanding each edge of F into a hyperedge containing it. Let Mk denote the matching of size k. Kang, Ni, and Shan [12] determined the Turán number of Berge-Mk. Our main result shows that if an r-uniform hypergraph H on n vertices has nearly as many edges as the extremal in their theorem without containing Berge-Mk, then H must be structurally close to certain well-specified graphs. Meanwhile, our result also implies several stability results, such as the stability version of the well-known Erdős-Gallai theorem (Erdős and Gallai, 1959 [5]).
Technological collaboration has become an important means of sharing technological risks and advancing key technological breakthroughs. However, existing studies often overlook patent value heterogeneity in technology representation and insufficiently integrate multiple technological drivers of collaboration, while also lacking effective modeling of their dynamic evolution. This study proposes a Dynamic Multi-View Graph Neural Network model based on Semantic and Quality fusion (SQ-DMVGNN). A semantic-quality fusion module is designed to project applicants’ textual semantic features and multidimensional quality attributes into a unified high-dimensional vector space, thereby capturing both technological semantics and patent quality. Then, the proposed framework constructs a dynamic multi-view graph structures to systematically decouple and integrate technological similarity, complementarity, and strategic trend, while modeling their evolution across time slices. Through the proposed framework, SQ-DMVGNN more accurately captures the multidimensional technological relationships underlying collaborations driven by high-quality technological resources, thereby improving both the performance and interpretability of collaboration prediction in dynamic environments. This study selects the representative semiconductor industry as empirical data, obtaining a total of 668,790 non-duplicated invention patents. Experimental results demonstrate that integrating patent quality into technological portfolios enables a more precise differentiation of technological heterogeneity among organizations. Compared with baseline methods grounded in different principles, the proposed model generally achieves near-optimal performance across major evaluation metrics, particularly in AUC, AP, MRR, and Hits@K. Ablation studies not only validate the effectiveness of each auxiliary view, temporal module, and patent quality, but also investigate model substitutions. Moreover, our work conducts parameter sensitivity analyses across different modules, verifying the stability of technological knowledge extraction and collaboration prediction.
One of the goals of 15-minute walkable neighborhoods is to provide equitable opportunities for all individuals within a 15-minute walk. This has motivated numerous studies to identify vulnerable groups living outside such neighborhoods using objective spatial data. However, very little research examines inequalities from a subjective perspective. More importantly, it remains unknown whether people outside such neighborhoods actually prefer them. Using data from a nationwide survey of 7,155 respondents across 81 Chinese cities, we employ multilevel linear regression models to assess inequalities in perceived accessibility, and multinomial logistic regression models to examine which population groups fall into a four‑group typology, i.e., HA‑HP (high accessibility & high preference), LA‑HP (low accessibility & high preference), LA‑LP (low accessibility & low preference), and HA‑LP (high accessibility & low preference), based on the match or mismatch between perceived accessibility and accessibility preference. The results indicate that socioeconomically disadvantaged groups (e.g., low‑education, low‑income, low‑car‑ownership, renters) tend to report poorer perceived accessibility to daily or recreational facilities. However, the typology analysis further shows that only some of them have unmet needs (LA‑HP), while others hold low accessibility preferences (LA‑LP). Moreover, different facility types exhibit distinct patterns. For daily facilities (e.g., shopping facilities, restaurants, bus stops), typical LA‑HP groups include low‑income, low‑education, and non‑full‑time employed residents. For recreational facilities (e.g., cultural facilities, sport facilities, leisure facilities), LA‑HP groups include men, low‑income groups, non‑local Hukou holders, and residents in low‑street‑density areas. These findings demonstrate that combining perceived accessibility with preference provides a more accurate and policy‑relevant identification of vulnerable groups than conventional approaches. Particular policy attention should be devoted to LA‑HP residents, i.e., those who desire better walkable access but perceive it as poor.
Gallium (Ga)-based biomaterials have emerged as versatile therapeutic platforms owing to the iron-mimicking properties and broad biological activities of gallium. Coordination engineering regulates Ga speciation, stability, transport, and therapeutic activation; however, optimizing individual physicochemical properties does not necessarily improve biological performance. Excessive coordination stability may restrict ligand exchange and therapeutic activation, whereas insufficient stability may cause premature dissociation, nonspecific exposure, and reduced target delivery.In this Review, we propose Biological Accessibility as a multidimensional, time-dependent property describing the spatiotemporal availability of biologically active gallium species to relevant biological environments, cells, and molecular or tissue targets throughout the therapeutic lifecycle. We first examine ligand chemistry, coordination thermodynamics, ligand-exchange kinetics, gallium speciation under physiological conditions, and disease-responsive coordination evolution. We then review representative coordination-engineered gallium platforms and their applications in antibacterial therapy, immunomodulation, cancer treatment, bone regeneration, bioimaging, and theranostics, highlighting coordination-regulated transport, localization, cellular accessibility, activation, retention, and clearance. Building on these advances, we develop a dynamic framework linking continuous coordination evolution with Biological Accessibility across systemic, pathological, and intracellular environments. We further discuss Therapeutic and Off-target Accessibility, disease-specific accessibility windows, mechanistic attribution of biological functions, quantitative evaluation, in vivo coordination evolution, predictive modeling, artificial intelligence-assisted design, biosafety, and clinical translation. This framework provides a conceptual basis for accessibility-guided design of next-generation gallium biomaterials and broader coordination-engineered therapeutic systems.
