high-resolution imaging

Scientific image annotation projects face cross-domain managerial challenges—including scarce data acquisition, inefficient resource allocation, inadequate annotator training, and pronounced human bias. To address these issues, this paper proposes the first general-purpose framework for preparing scientific image annotation projects. The framework systematically integrates objective definition, data availability assessment, multi-role team configuration, bias mitigation strategies, and an iterative annotator training mechanism, complemented by a recommended toolchain supporting integrated project management, quality control, and collaborative annotation. A novel closed-loop workflow—comprising bias detection, feedback integration, and retraining—is introduced to significantly enhance annotation consistency and efficiency. Empirical evaluation across multiple disciplines demonstrates that the framework reduces annotation costs by over 20%, improves project success rates, and strengthens knowledge base construction quality—thereby filling a critical research gap in standardized preparation guidelines for complex scientific image annotation.

Two-dimensional tissue sections struggle to reliably capture the spatial characteristics of local cellular interactions and rare cell populations within three-dimensional tissue architecture, while dense volumetric imaging remains prohibitively expensive. This study systematically evaluates, for the first time, the bias introduced by 2D sampling on local spatial statistics and proposes a geometry-aware sparse 3D reconstruction framework. By integrating phenotypic similarity with spatial proximity to associate cells across serial sections and incorporating cell type–specific shape priors, the method reconstructs high-fidelity single-cell 3D coordinates. Validated on both public imaging mass cytometry and in-house CODEX datasets, the approach significantly enhances the reliability of spatial analysis under limited imaging budgets, outperforming conventional 2D analyses. It enables structure-level 3D spatial resolution and provides quantitative guidance for experimental design regarding section spacing, coverage, and redundancy.

Large-scale high-resolution 3D volumetric data (e.g., X-ray tomography) continue to grow in size, posing challenges for existing segmentation and interactive analysis tools—including computational inefficiency, memory bottlenecks, and insufficient scalability for collaborative workflows. To address these, we propose a CUDA-based tiled execution architecture with strict memory management, integrated into the Annotat3D system via the Harpia library, enabling stable processing of ultra-large volumes under constrained single-GPU memory budgets. Our approach combines GPU-accelerated filtering, dynamic tiling scheduling, and human-in-the-loop interaction strategies to form an efficient 3D processing pipeline. Experimental evaluation demonstrates significant improvements over cuCIM and scikit-image in processing speed, memory efficiency, and strong scaling across multi-node HPC environments. The framework supports real-time, remote collaborative scientific imaging analysis.

This work addresses the challenge of synthesizing gigapixel images from ordinary photographs and sparse microscopic close-ups at extreme magnifications up to 350×, where preserving both fine material boundary details and large-scale structural consistency is difficult. The authors propose a two-stage cascaded generative framework: the first stage recovers global pattern coherence, while the second refines local textures, guided by segmentation masks to enable reference-driven synthesis in ambiguous regions. This approach achieves, for the first time, extreme-scale super-resolution tailored to everyday objects, generating gigapixel images that exhibit both microscopic material realism and macroscopic structural consistency on a newly curated dataset. Notably, it effectively maintains global coherence of repetitive geometric patterns—such as those in fabrics—and enables fully scalable, explorable visualization across all resolutions.

This work addresses the absence of a unified theoretical framework in computational imaging systems, which hinders effective diagnosis and optimization of reconstruction failures. The authors propose a universal grammar that decomposes any imaging forward model into a directed acyclic graph composed of eleven fundamental physical primitives. Building upon three root causes—information loss, carrier noise, and operator mismatch—they establish a tripartite decomposition theorem and a corresponding lifecycle gating mechanism to guide system design and calibration. Through graph-theoretic and information-theoretic structural decomposition, formal analysis of forward models, and cross-modal validation, the framework’s completeness and minimality are demonstrated across twelve imaging modalities spanning five carrier families, achieving empirical reconstruction performance gains of 0.8–13.9 dB.

This work addresses the challenge that critical scientific information in automated microscopy is often embedded in sequentially acquired spectral or functional response landscapes, which cannot be directly revealed by conventional imaging. To tackle this, the authors propose BEACON, a novel framework that introduces novelty-driven active exploration into microscope-based target space discovery for the first time. BEACON integrates deep kernel learning to dynamically model structure–response relationships and guides the system toward efficiently exploring diverse response regions. The study establishes a reproducible benchmarking protocol that explicitly disentangles exploration quality from optimization performance and introduces a metric for evaluating target space coverage. Experiments demonstrate that BEACON significantly outperforms classical acquisition strategies on offline datasets and has been successfully deployed on a scanning transmission electron microscope (STEM) for efficient real-time scientific discovery.

This work addresses the bottleneck in high-throughput video acquisition, where the readout and transmission bandwidth of a single chip struggles to keep pace with rapidly increasing pixel counts. The authors propose a compact, high-speed video camera architecture that, for the first time, integrates low-bit gradient imaging with a multi-scale convolutional neural network (CNN) for image reconstruction. This approach substantially reduces data volume while efficiently recovering high-resolution images. By leveraging the fast readout capability of gradient sensors and sub-micron pixel integration technology, the system effectively alleviates throughput and complexity constraints. Both simulations and real-world experiments demonstrate that the proposed method achieves excellent reconstruction quality while significantly enhancing system efficiency, enabling a compact single-chip design.

This work addresses the labor-intensive and expertise-dependent nature of computational imaging system design by proposing a method that automatically generates verifiable forward models from natural language instructions. Leveraging a formal specification language (spec.md) and a multi-agent architecture comprising Plan, Judge, and Execute modules, the approach combines a finite primitive basis to translate single-sentence descriptions into imaging systems with bounded reconstruction error. The study introduces a novel “design-to-reality error decomposition theorem,” which decouples total error into five independently controllable components, enabling cross-modal composition of high-dimensional (3D–5D) primitives. Evaluated across six real-world data modalities, the method achieves expert-level quality with 98.1 ± 4.2% fidelity and successfully produces ten novel imaging designs that surpass the capabilities of any single modality.

This work addresses the inherent ill-posedness and ambiguity of translational super-resolution at a single scale by proposing a multi-scale super-resolution method that requires no image priors. By leveraging low-resolution images acquired with incommensurate pixel pitches—such as those from different sensors or optical zoom settings—the approach constructs a well-posed reconstruction system. A stable solution is achieved through iterative least-squares optimization in the Fourier domain. Theoretical analysis demonstrates that incommensurate sampling guarantees the existence of a stable inverse for the system and reveals a fundamental trade-off between noise amplification and achievable reconstruction resolution. Experiments on both one-dimensional and two-dimensional scenarios successfully recover high-resolution images, validating the method’s efficacy and its practical potential for deployment on conventional imaging hardware, such as CCD sensors supporting pixel binning.

This study addresses the inefficiency and limited scalability of manually aligning functional diagrams—such as P&IDs—with 2D/3D as-built data in legacy industrial facilities lacking native digital models. To bridge this gap, the authors introduce IRIS-v2, the first publicly available multimodal industrial alignment dataset, which encompasses images, LiDAR point clouds, semantic segmentation masks, CAD models, 3D piping layouts, and corresponding P&ID schematics. Building upon this foundation, they propose an automated alignment method that integrates semantic segmentation, graph matching, and LiDAR point cloud processing. Evaluated in real-world industrial settings, the approach significantly reduces alignment time compared to manual workflows. This work establishes both a critical data resource and a practical pathway toward automating digital twin construction in complex industrial environments.