研究目的
To classify foodborne bacteria using hyperspectral microscope imaging technology coupled with convolutional neural networks for rapid detection and classification at the cellular level.
研究成果
The HMI technology coupled with CNN frameworks demonstrated potential as a novel optical diagnostic tool for the detection and classification of foodborne bacteria. U-Net performed well in generating accurate cell-ROI segmentation masks rapidly and automatically, while 1D-CNN achieved high classification accuracy in differentiation of spectral profiles.
研究不足
The manual extraction of spectral profile data using handcrafted segmentation methods is time consuming and cumbersome. Current chemometrics approaches with multifarious feature extraction protocols and lower classification accuracy also restrict the development of HMI technology.
1：Experimental Design and Method Selection:
Hyperspectral microscope imaging (HMI) technology combined with convolutional neural networks (CNN) was used for classifying foodborne bacterial species. U-Net and one-dimensional CNN (1D-CNN) were employed for cellular regions of interest (ROI) segmentation and spectral profile classification, respectively.
2：Sample Selection and Data Sources:
Purified isolates of five foodborne bacteria (Campylobacter fetus, Escherichia coli, Listeria innocua, Salmonella Typhimurium, and Staphylococcus aureus) were obtained and prepared for HMI data acquisition.
3：List of Experimental Equipment and Materials:
A self-assembled HMI system consisting of an upright microscope, an acousto-optic tunable filter (AOTF) spectrometer, an electron-multiplying charge-coupled device (EMCCD) camera, and a dark-field illuminator with a metal halide (MH) lamp.
4：Experimental Procedures and Operational Workflow:
Bacterial samples were prepared and scanned using the HMI system. Hypercubes were acquired in the wavelength range of 450–800 nm. Cell-ROI spectral data processing flow included segmentation, hypercube reconstruction, spectral profiles preparation, and dataset preparation.
5：Data Analysis Methods:
U-Net for segmentation and 1D-CNN for classification were compared with conventional methods (Otsu and Watershed for segmentation; PCA-KNN and PCA-SVM for classification). Performance was evaluated using mean pixel accuracy (MPA) and mean intersection over union (MIOU).
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
