There is a contradiction between the evolution rate of materials and the time resolution of SR-CT characterization in the in situ synchrotron radiation computed tomography(SR-CT)characterization of ultrafast evolution...There is a contradiction between the evolution rate of materials and the time resolution of SR-CT characterization in the in situ synchrotron radiation computed tomography(SR-CT)characterization of ultrafast evolution process.The sampling strategy of the ultra-sparse angle is an effective method for improving time resolution.Accurate reconstruction under sparse sampling conditions has always been a bottleneck problem.In recent years,convolutional neural networks have shown outstanding advantages in sparse-angle CT reconstruction given the development of deep learning.However,existing ideas did not consider the expression of high-frequency details in neural networks,limiting their application in accurate SR-CT characterization.A novel high-frequency information-constrained deep learning network(HFIC-Net)is proposed in response to this problem.Additional high-frequency information constraints are added to improve the accuracy of the reconstruction results.Further,a series of numerical reconstruction experiments are conducted to verify this new method,and the results indicate that the reconstruction results of HFIC-Net method effectively improve reconstruction quality.This new method uses only eight-angle projections to achieve the reconstruction effect of the filtered backprojection method(FBP)method in 360 projections.The results of the HFIC-Net method demonstrate clear boundaries and accurate detailed structures,correcting the misinformation caused by using other methods.For quantitative evaluation,the SSIM used to evaluate image structure similarity is increased from 0.1951,0.9212,and 0.9308 for FBP,FBP-Conv,and DDC-Net,respectively,to 0.9620 for HFIC-Net.Finally,the results of actual SR-CT experimental data indicate that the new method can suppress artifacts and achieve accurate reconstruction,and it is suitable for the in situ SR-CT accurate characterization of ultxafast evolution process.展开更多
Computational fluorescence microscopy constantly breaks through imaging performance through advanced opticalmodulation technologies;however, conventional theoretical modeling and experimental measurement approachesare...Computational fluorescence microscopy constantly breaks through imaging performance through advanced opticalmodulation technologies;however, conventional theoretical modeling and experimental measurement approachesare challenging to meet the demand for accurate system characterization of diverse modulations. To this end, wepropose a point spread function (PSF) decoupling method that is intrinsically compatible with the optimaldemodulation in computational microscopic imaging modality. The critical core lies in designing a sample prior-basedcomputational imaging strategy, in which a regular fluorescent sample instead of generally used sub-diffractionlimited particles acts as a system modulator to demodulate the system response. PSF consequently can becomputationally optimized through the strong support from the modulated sample prior, achieving accurate nonparametricsystem characterization and thereby avoiding the modeling difficulty and the low signal-to-noise ratiomeasurement errors of the system specificity. Experimental results across various biological tissues demonstrated andverified that the proposed PSF decoupling method enables excellent volumetric imaging comparable to confocalmicroscopy and multicolor, large depth-of-field imaging under aperture modulation. It provides a promisingmechanism of system characterization and computational demodulation for high-contrast and high-resolutionimaging of cellular and subcellular biological structures and life activities.展开更多
基金supported by the National Nature Science Foundation of China(Nos.12027901 and 12041202)Synchrotron Radiation Joint Fund of University of Science and Technology of China(Nos.KY2090000059 and KY2090000054)。
文摘There is a contradiction between the evolution rate of materials and the time resolution of SR-CT characterization in the in situ synchrotron radiation computed tomography(SR-CT)characterization of ultrafast evolution process.The sampling strategy of the ultra-sparse angle is an effective method for improving time resolution.Accurate reconstruction under sparse sampling conditions has always been a bottleneck problem.In recent years,convolutional neural networks have shown outstanding advantages in sparse-angle CT reconstruction given the development of deep learning.However,existing ideas did not consider the expression of high-frequency details in neural networks,limiting their application in accurate SR-CT characterization.A novel high-frequency information-constrained deep learning network(HFIC-Net)is proposed in response to this problem.Additional high-frequency information constraints are added to improve the accuracy of the reconstruction results.Further,a series of numerical reconstruction experiments are conducted to verify this new method,and the results indicate that the reconstruction results of HFIC-Net method effectively improve reconstruction quality.This new method uses only eight-angle projections to achieve the reconstruction effect of the filtered backprojection method(FBP)method in 360 projections.The results of the HFIC-Net method demonstrate clear boundaries and accurate detailed structures,correcting the misinformation caused by using other methods.For quantitative evaluation,the SSIM used to evaluate image structure similarity is increased from 0.1951,0.9212,and 0.9308 for FBP,FBP-Conv,and DDC-Net,respectively,to 0.9620 for HFIC-Net.Finally,the results of actual SR-CT experimental data indicate that the new method can suppress artifacts and achieve accurate reconstruction,and it is suitable for the in situ SR-CT accurate characterization of ultxafast evolution process.
基金supported by grants from the National Natural Science Foundation of China(NSFC)(62275173,62175109,62371311)Shenzhen Fundamental Research Program(JCYJ20220531101204010)+1 种基金Shenzhen Higher Education Stable Support Program(20231122025852001)Scientific Instrument Developing Project of Shenzhen University(2023YQ009).
文摘Computational fluorescence microscopy constantly breaks through imaging performance through advanced opticalmodulation technologies;however, conventional theoretical modeling and experimental measurement approachesare challenging to meet the demand for accurate system characterization of diverse modulations. To this end, wepropose a point spread function (PSF) decoupling method that is intrinsically compatible with the optimaldemodulation in computational microscopic imaging modality. The critical core lies in designing a sample prior-basedcomputational imaging strategy, in which a regular fluorescent sample instead of generally used sub-diffractionlimited particles acts as a system modulator to demodulate the system response. PSF consequently can becomputationally optimized through the strong support from the modulated sample prior, achieving accurate nonparametricsystem characterization and thereby avoiding the modeling difficulty and the low signal-to-noise ratiomeasurement errors of the system specificity. Experimental results across various biological tissues demonstrated andverified that the proposed PSF decoupling method enables excellent volumetric imaging comparable to confocalmicroscopy and multicolor, large depth-of-field imaging under aperture modulation. It provides a promisingmechanism of system characterization and computational demodulation for high-contrast and high-resolutionimaging of cellular and subcellular biological structures and life activities.
