Tong Wang

Excitation–emission matrix (EEM) fluorescence spectroscopy is widely used for characterizing complex samples in food analysis and environmental monitoring, but its practical deployment is limited by the trade-off between spectral resolution and acquisition speed. High-resolution EEM acquisition is time-consuming, whereas rapid low-resolution measurements lose analytically important spectral details. Herein, we introduce SuperEEM, a chemically constrained super-resolution framework for converting rapidly acquired sparse EEMs into analytically reliable high-resolution EEMs. Unlike generic image super-resolution approaches that mainly optimize visual similarity, SuperEEM couples an EEM-tailored U-Net–GAN architecture with a quantitative-fidelity-oriented loss function to preserve fluorescence intensity, peak topology, and excitation–emission relationships. The analytical reliability of SuperEEM was validated using three representative tasks: polycyclic aromatic hydrocarbon (PAH) quantification, edible oil classification, and beer classification. Compared with LR EEMs, SuperEEM-reconstructed EEMs substantially improved downstream performance, achieving PAH quantification comparable to original HR EEMs and improving classification accuracy from 77.4% to 90.3% for edible oils and from 73.3% to 86.7% for beers. Under the tested instrumental settings, SuperEEM enabled a 14-fold improvement in EEM acquisition efficiency while maintaining analytical reliability. A graphical user interface (GUI) was further developed to support model application, visualization, and result export. These results demonstrate the potential of SuperEEM for rapid, high-throughput fluorescence analysis.

Rapid and reliable authentication of Ziziphi Spinosae Semen (ZSS) is essential for quality control because visually similar adulterants are difficult to distinguish using conventional inspection methods. In this study, hyperspectral imaging (HSI) was integrated with a dual-branch spectral–spatial attention network (HSSANet) for the non-destructive authentication of ZSS and its adulterants. The framework simultaneously exploits complementary spectral and spatial information: a lightweight one-dimensional convolutional neural network extracts wavelength-dependent spectral representations from seed-level mean spectra, whereas a ResNet18-based branch captures morphological and textural features from PCA-reduced hyperspectral images. A cross spectral–spatial attention module was introduced to use spatial contextual information to adaptively recalibrate spectral representations prior to feature fusion. Compared with conventional spectral classifiers and representative convolutional neural networks, HSSANet achieved the best overall classification performance, with external test accuracies of 97.77% for binary authentication and 96.59% for three-class discrimination. Spectral attribution analysis and gradient-weighted class activation mapping further identified the wavelength regions and spatial areas contributing to model decisions, providing complementary interpretation of the learned spectral–spatial representations. An automated graphical interface was additionally implemented to demonstrate the practical applicability of the proposed workflow. These results indicate that adaptive integration of spectral and spatial information provides an effective and interpretable analytical strategy for the rapid authentication of morphologically similar herbal materials.

Mineral oil contamination in vegetable oils poses a serious threat to food safety and consumer health. In this study, we reported an on-site compatible analytical strategy based on laser light scattering-enhanced deep computer vision for detecting mineral oil contamination in vegetable oils. The strategy integrates saponification-induced phase and turbidity contrast with laser-enhanced scattering visualization to transform trace mineral oil into visually discriminative signals. To accurately analyze these signals, we proposed a novel, lightweight, and efficient deep learning model (Oil-MobileNet). After optimizing chemical reaction conditions, the effects of three illumination sources (green laser, red laser, and laser-free) on image acquisition were systematically examined. Subsequently, Oil-MobileNet was evaluated on binary and multiclass classification. Comparative analyses with four baseline models demonstrated that the combination of green laser illumination and Oil-MobileNet achieved the best classification performance, enabling reliable discrimination of mineral oil contamination down to 0.05% (v/v) under the defined operational criteria. This practical detection capability outperformed human visual inspection with green laser (0.1%) and commercially available saponification-based kits (0.9-3%). The contaminated level prediction models were also established using these architectures. In addition, interpretability studies were conducted to elucidate the model's decision-making mechanism. Finally, the well-trained models were deployed in a user-friendly graphical user interface for the accurate determination of mineral oil contamination.

LLM-based visual inspection framework with two paths: text-only inference and vision-language inference. A lightweight GUI enables one-click automation. Models are evaluated on ripeness, freshness, and authenticity tasks.

Aflatoxin B1 (AFB1) is a hazardous mycotoxin commonly found in agricultural products, necessitating sensitive and reliable detection methods. In this work, a novel aptamer-based fluorescence sensing strategy was developed, in which DNAzyme acts as both signal generator and amplifier. In the absence of AFB1, a blocking strand hybridizes with the DNAzyme arm, inhibiting its catalytic activity. When AFB1 is present, the aptamer specifically binds the target, releasing the blocking strand and restoring DNAzyme activity. The activated DNAzyme cleaves a fluorophore–quencher labeled substrate, leading to enhanced fluorescence. To achieve accurate quantification, a probe technique-based generalized multivariate standard addition strategy (GMSAprobe) was employed. The method enabled reliable determination of AFB1 in peanut samples with recoveries of 96-102% and a detection limit of 0.56 μg/kg, consistent with results from a commercial ELISA kit and well below regulatory limits.

细胞是生物体的基本结构和功能单位，其形态和状态能够直接反映机体的生理与病理变化，因此细胞学研究在疾病诊断中具有重要意义。由于细胞结构复杂多样，实现高精度的细胞识别与分析一直是该领域的核心挑战。传统的细胞图像分析方法往往存在效率低、精度不足和成本高等问题。随着第四次工业革命的推进，人工智能的快速发展为细胞学研究带来了新的契机。人工智能与单细胞显微图像的结合显著提升了细胞识别的准确率与分析效率。本综述基于人工智能单细胞显微图像分析进展，对以下应用方向进行概述：细胞图像重建、细胞图像分割、细胞图像分类、细胞图像翻译、多模态整合以及动态与功能分析。最后，探讨了当前细胞图像分析面临的挑战并展望未来研究方向。