Coronary artery disease (CAD), primarily caused by arterial stenosis or occlusion, remains a leading cause of mortality worldwide. Invasive coronary angiography (ICA) is the clinical gold standard for CAD diagnosis and treatment planning, yet its effectiveness critically depends on accurate coronary artery segmentation and stenosis assessment. However, ICA images are often affected by low contrast, severe noise, and complex multi-scale vascular morphology, substantially challenges existing segmentation methods. This study aims to improve the robustness and reliability of ICA-based coronary artery segmentation and stenosis detection by learning structure-consistent representations that jointly capture geometric priors, cross-dataset variability, long-range spatial dependencies, and frequency-aware vascular features. We propose SFD-Mamba2Net, an end-to-end framework for coronary artery segmentation and stenosis detection. A Curvature-Aware Structural Enhancement (CASE) module is introduced in the encoder to highlight elongated vascular structures while suppressing background interference. To alleviate performance degradation induced by inter-dataset discrepancies, a Supervised Domain Adaptation (SDA) strategy is incorporated at the bottleneck to encourage domain-invariant feature learning. Long-range spatial dependencies and complex vessel topology are modeled via an Axial-Alternating Dual-Stream Mamba2 (AA-DS Mamba2) module. In the decoder, a Progressive High-Frequency Perception (PHFP) module integrates multi-level wavelet decomposition with deep convolution to refine high-frequency vascular details while preserving global structural. Experiments on three large-scale ICA datasets show that SFD-Mamba2Net achieves superior segmentation accuracy and strong cross-domain generalization, obtaining Dice scores of 89.31%, 77.01%, and 74.14% on JSPH-ICA, DCA1, and ARCADE, respectively. It also achieves high true positive and positive predictive values in stenosis detection, demonstrating robustness and clinical potential.
The rapid advancement of artificial intelligence is paving the way for shared autonomous vehicles (SAVs), which are expected to replace private vehicles and enhance transportation efficiency. To enable the continuous operation of SAVs, Dynamic Wireless Charging (DWC) technology has emerged as an ideal solution, allowing SAVs to charge while in motion. However, conventional vehicle routing algorithms are inadequate for SAV pickup, delivery, and charging problems, lacking flexibility under dynamic traffic conditions and failing to incorporate DWC. To address these challenges, we introduce a Two-Stage Deep Reinforcement Learning (TDRL) framework. This framework develops an initial current-state route construction strategy that considers route cost, state of charge (SOC), and battery health, together with a traffic-responsive route-replanning strategy that updates the remaining executable route when traffic conditions evolve during operation. The first stage of TDRL employs a novel Heterogeneous Multi-Layer Perceptron Attention (HMLP-Att) mechanism to integrate diverse node, vehicle, and traffic-state information for DWC node selection and global route construction at the initial decision epoch. The second stage starts from this initial executable route and re-embeds the updated traffic, vehicle, passenger, and visitation states, enabling strategic route adjustments through the deletion and reinsertion of unvisited node pairs under a specially designed masking scheme. Experimental results demonstrate that TDRL outperforms existing heuristic and advanced DRL algorithms in reducing route cost, maintaining SOC stability, and reducing estimated route-induced SOH degradation across different scales. Moreover, compared with the initial TDRL-1 route constructed under the same departure traffic state, TDRL-2 significantly mitigates the impact of subsequent traffic fluctuations on route cost, with reductions of 16.2% for the 50-node model and 23.7% for the 100-node model, while also improving SOC distribution. The specially designed masking scheme further exhibits good robustness across different traffic peak periods. Finally, the sensitivity analyses of SAV routes and DWC road usage relative to traffic congestion are presented.
This paper addresses the quasilinear Schr & ouml;dinger equation with critical growth in the zero-mass case: -Delta u - u Delta(u(2)) = lambda g(x)|u|(p-2)u + |u|(2 & sdot;2 & lowast;-2)u, x is an element of & Ropf;(N) , where N >= 3, lambda > 0, p is an element of (4, 2 & sdot;2(& lowast;)) or p is an element of (1, 2) with 2(& lowast;) = 2N/N-2 the critical Sobolev exponent, and g is an element of L-p0 (& Ropf;(N)), where p(0) = 2.2 & lowast; 2 & sdot; 2 & lowast; - p if p is an element of (4, 2 & sdot;2(& lowast;)), p0 = 2(& lowast;)/2 & lowast;- p if p is an element of (1, 2). To address the inherent lack of compactness in the zero-mass case, we develop a perturbative variational framework inspired by established techniques. For the case p is an element of (4, 2 & sdot; 2(& lowast;)), we establish the existence of a positive-energy ground state solution when lambda is sufficiently large. For the case p is an element of (1, 2), we prove that for small lambda > 0, the equation admits a ground state solution with negative energy. Additionally, for p is an element of (1, 2), we show the multiplicity of solutions to the above equation for small lambda
Oracle Bone Inscription (OBI) recognition is important for the digital analysis and study of ancient Chinese writing. However, automatic recognition remains challenging because OBI characters ex hibit substantial structural variation, degradation, and fine-grained inter-class similarity. Although deep learning has advanced general image recognition, existing attention mechanisms may still provide limited benefits for capturing subtle OBI-specific variations. Recent layer-attention meth ods enhance cross-layer interaction by aggregating representations from preceding layers, but their improvements in OBI recognition remain modest. To address this issue, we propose Multi-Scale Layer Attention (MSLA), which integrates multi-scale token enrichment with progressive cross-layer attention. MSLA introduces global and local representations into cross-layer interaction, enabling richer feature aggregation across layers. Extensive experiments on four OBI benchmarks, including Oracle-MNIST, OBC306, HUST-OBS, and EVOBC, show that MSLA consistently improves recognition performance over the evaluated attention baselines under controlled settings. In par ticular, on OBC306 with ResNet-20, MSLA improves the plain backbone by 3.56 percentage points and outperforms the strongest competing attention method, MRLA-L, by 0.79 percentage points, while introducing moderate computational overhead.
This study combines machine learning, fragment-based molecular generation, and synthetic-accessibility analysis to identify promising small molecule acceptors (SMAs) for organic solar cells. Twelve regression algorithms were assessed using 60% training, 20% validation, and 20% test partitions. Gradient Boosting showed the strongest generalization, with a test R² of approximately 0.84–0.85. BRICS decomposition of 1027 parent molecules produced 3779 unique fragments, which were recombined to generate 50,000 unique, chemically valid candidates. The trained model predicted electron reorganization energies for the generated library, while the synthetic accessibility (SA) score was used to estimate synthetic feasibility. Chemical-space analysis using principal component analysis and t-distributed stochastic neighbor embedding confirmed broad structural diversity. The 1000 candidates with the lowest predicted reorganization energies were retained, after which 40 structures with the most favorable SA scores were prioritized. Similarity-matrix, hierarchical-clustering, t-SNE, and network analyses showed that the selected candidates span several structural families.
