Researches On Fast And High-Resolution Raman Imaging Technology And High-Throughput Spectral Analysis Methods

Raman spectroscopy is a non-invasive,label-free,highly sensitive tool to probe chemical composition,while Raman imaging is a functional imaging method which can characterize the spatial distribution of specific chemical information.Hyperspectral Ra-man microscopy plays an important role in research in a wide application space,includ-ing in the life sciences where it can probe cell metabolism,perform tissue histopathology(such as identifying cancerous tissues),among many others.The thesis describes sev-eral instrumentation and analysis improvements to standard Raman imaging systems to achieve faster,higher resolution,and more accurate spectral image data.The key points are fast confocal point-scanning Raman imaging,superresolution hyperspectral Raman imaging and high-throughtput spectral analysis method suitable for for large Raman datasets.The detailed research contents and achievements are as follows:1.In view of the slow imaging speed of traditional confocal laser-scanning Raman imaging system,the thesis proposes a fast,context-aware compressive Raman imaging strategy based on the known spatial distribution of the sample obtained from bright-field imaging.Concrete implementation procedures of the proposed method are:using a priori spatial locations of samples provided by a bright-field imaging channel,information-rich and "blank" regions on sample plane are segmented,then information-rich regions are imaged through compressive sens-ing Raman imaging,finally reconstructing hyperspectral Raman images using an optimized compressive sensing reconstruction algorithm,therefore further im-proving the imaging speed of standard confocal laser-scanning Raman imaging systems.The thesis utilizes the proposed method to image standard 1?m mi-crospheres and fission yeast cells,and the imaging results imply that compared to classical confocal laser-scanning Raman imaging,fast context-aware Raman compressive imaging can achieve 5-to 10-fold faster imaging speed,and also keep imaging quality highly closer to point-scanning imaging.2.In order to realize truly two-dimensional superresolution hyperspectral Raman imaging,the thesis proposed a multifocal line-scanning superresolution Raman imaging approach based on galvo-modulated virtually structured illumination.The implementation procedures of the proposed imaging method are:first,line-shaped,sinusoidally-modulated structured illumination patterns are generated with random illumination orientations on the sample plane by two-dimensional galvo-mirrors during a Raman measurement exposure time.The Raman emission is then de-scanned by the galvos and a third galvo-mirror re-scans the line-shaped Raman emission onto the narrow slit of Raman spectrograph.In this way,lines of arbitrary orientations can be imaged onto a fixed spectrograph slit.Finally,super-resolved images are reconstructed using optimized SIM algorithms,e.g.Wiener-SIM or TV-SIM.The thesis demonstrates the feasibility and performance of the proposed imaging method on standard nanospheres and some preliminary results have been achieved,showing that the proposed two-dimensional superres-olution hyperspectral Raman imaging based on virtually structured illumination can break the optical diffraction limit with theoretical 1.6-fold improvement in spatial resolution.3.Finally,the thesis explores the utility of convolutional neural networks(CNNs)for recovering concentrations of individual chemicals from mixed Raman spectra.A problem with current CNNs used for concentration analysis are that they cannot predict concentrations of multiple spectral components simultaneously.The the-sis thus proposed an optimized multi-prediction convolutional neural regression network for concentration qualification of mixed spectra.Detailed implementa-tion procedures of the proposed method are:the thesis proposes and designs an optimized CNN model based on the published spectral concentration prediction CNN models,firstly simulate and generate artificial Raman spectra of mixtures as training dataset of the proposed CNN model,then use reasonable training hyper-parameters to train a highly sensitive and efficient multi-concentration prediction CNN model for sub-component qualifications from mixed Raman spectra.The thesis uses measured Raman spectra of mixtures to verify the prediction accu-racy and efficiency of the trained multi-prediction convolutional neural regres-sion network.The thesis shows that the prediction results using the proposed multi-prediction CNN model are consistent with the current state-of-the-art tra-ditional method(Asymmetric Least Squares(AsLS)fitting),yet achieves these results about 45-fold faster than AsLS.Thus,it is highly appropriate for large datasets,such as those generated in this thesis,where>1,000,000 spectra must be independently analyzed.