Reconstructing in vivo spatially offset Raman spectroscopy of human skin tissue using a GPU-accelerated Monte Carlo platform

As one type of spatially offset Raman spectroscopy(SORS), inverse SORS is particularly suited to in vivo biomedical measurements due to its ring-shaped illumination scheme. To explain inhomogeneous Raman scattering during in vivo inverse SORS measurements, the light–tissue interactions when excitation and regenerated Raman photons propagate in skin tissue were studied using Monte Carlo simulation. An eight-layered skin model was first built based on the latest transmission parameters. Then, an open-source platform, Monte Carlo e Xtreme(MCX), was adapted to study the distribution of 785 nm excitation photons inside the model with an inverse spatially shifted annular beam. The excitation photons were converted to emission photons by an inverse distribution method based on excitation flux with spatial offsets Δs of 1 mm, 2 mm, 3 mm and 5 mm. The intrinsic Raman spectra from separated skin layers were measured by continuous linear scanning to improve the simulation accuracy. The obtained results explain why the spectral detection depth gradually increases with increasing spatial offset, and address how the intrinsic Raman spectrum from deep skin layers is distorted by the reabsorption and scattering of the superficial tissue constituents. Meanwhile, it is demonstrated that the spectral contribution from subcutaneous fat will be improved when the offset increases to 5 mm, and the highest detection efficiency for dermal layer spectral detection could be achieved when Δs = 2 mm. Reasonably good matching between the calculated spectrum and the measured in vivo inverse SORS was achieved, thus demonstrating great utility of our modeling method and an approach to help understand the clinical measurements.

基于纳米散射结构的可集成光学神经网络及其逆向设计

基于集成光学和硅光子学的光学神经网络硬件有很多优势:集成度高、速度快并且与CMOS工艺兼容.然而,目前的集成光学神经网络尺寸较大,很难扩展到大量神经元(>1000),实现大规模计算.本文提出了一种基于光学散射单元的神经网络硬件架构,除了具备一般光学神经网络的优势外,突出的优势是尺寸小,易于大规模扩展.光学散射单元允许光在一个小区域中发生散射,通过逆向设计散射区域结构,实现目标的计算功能.光学散射单元在一个很小的尺寸下,提供了很大的优化自由度,研究表明要实现一个4 4的矩阵乘法,计算单元尺寸只需要4 4μm^2.基于光学散射单元,本文设计了光学神经网络,在经典图像识别测试集MNIST上实现了97.1%的准确度.此外,这种光学散射单元还可以适用于相干光和非相干光.本研究提供了一个新的光学神经网络架构,能在不影响效率和功能下减小神经网络硬件尺寸.