AI for Chemical and Biomedical Imaging: From Data Acquisition to Actionable Insight Authors: Yong Fan, Sokratis Makrogiannis, Rong Zhou, Peng Chen Published: 2026 Chemical and biomedical imaging are entering an era in which the central challenge is no longer simply to acquire images, but to extract reliable, quantitative, and actionable information from increasingly complex data. Contemporary imaging platforms capture spatial, temporal, spectral, molecular, and functional information across scales ranging from single molecules and intracellular organelles to tissues, organs, and whole-brain networks. However, the acquisition and interpretation of these rich datasets are frequently complicated by noise, limited spatial or temporal resolution, incomplete sampling, motion, radiation exposure, lengthy acquisition, and the need for substantial expertise. Artificial intelligence (AI) provides a powerful framework for addressing these challenges. AI methods can improve image acquisition, reconstruction, enhancement, segmentation, classification, and quantitative analysis. When integrated closely with imaging instrumentation and domain knowledge, they can also support adaptive acquisition, real-time interpretation, and ultimately closed-loop experimental or process control. The Special Issue “AI for Chemical and Biomedical Imaging” was assembled to highlight advances at this rapidly evolving interface and to illustrate how AI is reshaping both chemical and biomedical imaging. The contributions span microscopy, magnetic resonance imaging (MRI), ultrasound, optical coherence tomography (OCT), nuclear imaging, cryogenic transmission electron microscopy (cryo-TEM), hyperspectral imaging, functional MRI, single-particle tracking, and imaging-guided manufacturing. Collectively, they demonstrate that AI is not merely a downstream analysis tool. Rather, it can contribute throughout the imaging pipeline by reducing acquisition requirements, reconstructing and enhancing images, identifying structures and abnormalities, inferring molecular dynamics, integrating complementary modalities, and converting complex imaging data into information that can guide scientific and clinical decisions. Three reviews provide broad conceptual and methodological foundations for the Special Issue. Zhu et al. examine supervised, unsupervised, and self-supervised approaches for mapping functional brain organization, including individualized functional parcellation and validation through test–retest reproducibility, functional homogeneity, task-based fMRI, electrophysiology, and cross-modal comparison (DOI: 10.1021/cbmi.5c00092). Zhang et al. review machine learning across the single-particle tracking workflow, from particle detection and trajectory linking to motion classification, denoising, uncertainty quantification, and biophysical inference (DOI: 10.1021/cbmi.5c00146). Yu et al. discuss the convergence of AI with coherent Raman scattering microscopy, highlighting spectral analysis, image enhancement, virtual staining, and automated histopathological assessment as pathways toward clinical translation (DOI: 10.1021/cbmi.5c00199). Across these areas, the reviews identify shared priorities, including representative training data, generalizability, biological validation, uncertainty estimation, and an appropriate balance between predictive performance and physical interpretability. They also point toward multimodal integration, foundation and self-supervised models, simulation-to-reality transfer, federated learning, and physics-informed and explainable AI. Together, they reinforce a central theme of this Special Issue: the value of AI depends not only on predictive accuracy, but also on the extent to which models incorporate imaging physics, preserve biologically meaningful information, quantify uncertainty, and generalize beyond the data used for their development. Several contributions demonstrate how AI can obtain more information from fewer, faster, lower-dose, or complementary measurements. Li et al. use a transformer-based approach to correct B0 inhomogeneity from reduced Z-spectrum sampling in GluCEST and NOE MRI, illustrating how learning-based methods can shorten acquisition while preserving quantitative information (DOI: 10.1021/cbmi.5c00071). Taragola et al. investigate cross-modality transfer from PET for denoising low-count μSPECT images when target-domain training data are limited (DOI: 10.1021/cbmi.5c00270). Song et al. use preoperative MRI as a source of anatomical prior knowledge for three-dimensional freehand ultrasound reconstruction, reducing accumulated drift and supporting more reliable MRI–ultrasound fusion (DOI: 10.1021/cbmi.5c00076). These studies illustrate complementary forms of data efficiency. Learned sampling can reduce acquisition requirements, transfer learning can improve image quality in data-limited domains, and multimodal information can constrain otherwise underdetermined reconstruction problems. More broadly, they show how AI can help redesign acquisition and reconstruction rather than operating solely on images after acquisition. They also provide a practical reminder that increasing architectural complexity does not necessarily improve performance, particularly when robustness, computational efficiency, and limited training data are important considerations. Automated segmentation constitutes another major theme. Liu et al. combine transformer-based global contextual modeling with convolutional extraction of local features for efficient three-dimensional biomedical image segmentation (DOI: 10.1021/cbmi.4c00102). Fang et al. develop a two-stage localization and semantic segmentation workflow to quantify Schlemm’s canal in visible-light OCT images acquired in vivo from mouse eyes (DOI: 10.1021/cbmi.5c00156). Michael et al. combine holotomographic microscopy with deep learning to segment endothelial cell mitochondria without fluorescent labeling, enabling analysis in a more native cellular state (DOI: 10.1021/cbmi.4c00100). Together, these studies show that segmentation is not a generic or modality-independent task. Effective solutions must account for image dimensionality, target geometry, contrast mechanism, computational requirements, and the intended use of their outputs. AI is also extending imaging from structure delineation toward quantitative characterization, classification, and quality control. Gutierrez and Robinson develop an automated workflow for analyzing cryo-TEM images of adeno-associated virus capsids, combining image segmentation and capsid classification with an interface that reduces the burden of active labeling (DOI: 10.1021/cbmi.5c00070). Kumar et al. use an unsupervised generative framework with spectral and spatial attention to analyze cholangiocarcinoma in hyperspectral images without requiring extensive manual annotation (DOI: 10.1021/cbmi.5c00054). Although these studies address different applications, both demonstrate that automation must be evaluated in relation to the intended use of its outputs. In therapeutic product quality control, improved efficiency must be accompanied by transparency, interpretable image features, and corroboration using complementary analytical measurements. In hyperspectral pathology, unsupervised learning may help exploit high-dimensional chemical information when specialist annotations are scarce or costly. For translation, however, performance should remain anchored to expert-provided references and independently validated data. The value of automation lies not only in greater throughput, but also in improving the consistency, traceability, and reliability of imaging-derived measurements. Two contributions focus on extracting molecular transport information using complementary computational strategies. Fan et al. introduce D-Blur, which uses information encoded in motion-blurred images to localize particles and estimate diffusion coefficients without explicit trajectory linking (DOI: 10.1021/cbmi.5c00042). Chatterjee et al. use physically meaningful trajectory descriptors, systematic feature selection, and optimized tree-based classifiers to characterize heterogeneous motion in three-dimensional single-particle tracking data (DOI: 10.1021/cbmi.5c00057). These studies illustrate the continuing value of methodological diversity in AI-enabled imaging. End-to-end deep learning can extract information that is difficult to recover using conventional analysis, whereas feature-based machine learning can retain physical interpretability and computational efficiency. The most appropriate strategy depends on the image-formation process, the available data, the desired level of interpretability, and the scientific question being addressed. Extending AI from image analysis toward real-time process monitoring, Garcia et al. extend AI-enabled segmentation to process monitoring in three-dimensional ice printing (DOI: 10.1021/cbmi.5c00064). By using an optical-flow method to generate training data for a faster neural network, the study establishes a foundation for near-real-time monitoring of transparent and optically complex structures. This approach illustrates how a computationally intensive analytical procedure can generate training data for a model capable of rapid operation. It may support future closed-loop control and has broader relevance to transparent-material manufacturing, microfluidics, organ-on-chip systems, tissue engineering, and biomanufacturing. Taken together, the contributions reveal several important transitions in AI-enabled imaging. AI is moving from retrospective analysis toward acquisition-aware measurement, as illustrated by reduced sampling, low-count denoising, multimodal reconstruction, and near-real-time process monitoring. Information is increasingly being transferred across modalities and data domains, allowing knowledge obtained in one imaging setting to improve reconstruction or analysis in another. Dependence on exhaustive manual annotation is also being reduced through unsupervised and self-supervised learning, transfer learning, and algorithmically generated training data, while foundation models represent a promising direction for further reducing reliance on task-specific annotations. Carefully selected handcrafted features remain valuable for efficient and interpretable modeling. Most importantly, AI helps to transform images into quantitative descriptions of chemical composition, tissue morphology, physiological response, functional brain organization, molecular transport, and manufacturing dynamics. This transition from image interpretation to quantitative inference is central to the scientific and translational value of the field. Looking ahead, progress in AI-enabled imaging will require closer integration of algorithms, instruments, and domain knowledge. Important priorities include foundation and self-supervised models, multimodal integration, transfer learning across modalities and domains, and physics-informed, uncertainty-aware, and explainable AI. At the same time, performance on a held-out test data set is not sufficient to establish scientific, clinical, or industrial utility. Models must be evaluated for generalizability, robustness, calibration, and failure modes across instruments, acquisition protocols, specimens, populations, and institutions. Representative data, external and prospective validation, transparent reporting, standardized evaluation, and continued expert oversight will be essential for developing AI methods that are not only accurate, but also reliable and scientifically meaningful. We thank all the authors for their outstanding contributions to this Special Issue and for demonstrating the breadth of AI applications in chemical and biomedical imaging. We also thank the reviewers for their rigorous and timely evaluations and the editorial and production teams of Chemical & Biomedical Imaging for their dedication and support. We hope that this collection will stimulate collaboration across disciplines and encourage the development of AI methods that are not only accurate and efficient, but also interpretable, reproducible, physically and biologically grounded, and aligned with meaningful scientific and clinical needs.