Rapid Wavefront Reconstruction And Compensation For Biological Tissue Based On Machine Learning

Brain science is considered to be one of the most important research fields in the21 st century.The structure and function of neural circuits are currently the focus and difficulty of global brain science research.High-resolution imaging(brain reading)and precise optogenetic light control(brain writing)deep inside the brain in a non-invasive manner will bring a milestone breakthrough in neural circuit research.However,brain tissue is an opaque and strong scattering medium.Optical distortion and scattering occur when light is transmitted inside it.The resulting wavefront aberration limits the focusing accuracy of the optical system and seriously affects the imaging resolution and light control precision deep into the brain tissue.Adaptive optics is one of the effective technologies of wavefront aberration detection and correction.However,traditional adaptive optics requires several seconds to tens of minutes to reconstruct the wavefront,which makes it difficult to apply deep light focusing in living biological tissues,which restricts deep penetration highresolution imaging and precise optogenetic light control.Therefore,how to quickly achieve wavefront reconstruction and compensation of biological tissues and achieve high-quality light focusing deep in biological tissues is one of the key problems to be solved in this field.In response to the above problems,this paper proposes a fast indirect detection and compensation method for wavefront aberration based on machine learning.We design a supervised learning framework,and establish the mapping relationship between the intensity distribution of distorted focal spots and the wavefront aberration(characterized by Zernike coefficient)for the first time.This method realizes the rapid reconstruction of the wavefront with only a single intensity image.The experiment results show that for 1mm-thick artificial tissue slices and 300?m-thick mouse brain slice samples,the method achieves wavefront aberration detection and compensation within 200 ms,greatly improving the beam focusing accuracy.Microscopic imaging experiments show that this method significantly improves the imaging clarity of 150?m mouse brain tissue slices,and optogenetic light regulation experiments demonstrate the practical feasibility and practical value of this method.Under the same number of optical modes,the detection speed of this method is 2-3 orders of magnitude higher than that of traditional adaptive optics technology,which provides a new idea for realizing high-resolution optical imaging and precise light control deep in living biological tissues.Although the convolutional neural network model achieves good high-speed wavefront reconstruction and compensation effects,the training time is considerably long because of the massive training data,which will affect the research or observation for new sample.In response to this practical problem,this paper proposes a method for predicting wavefront aberrations in biological tissues based on extreme learning.We use extreme learning to give a closed-form solution to the mapping between wavefront and light intensity distribution to achieve fast training,which further simplifies the prediction process.The experimental results show that for artificial tissue slices,the model training speed is increased by about 13.4 times while achieving the same precision level as the convolutional neural network.In addition,we find that the wavefront aberration measurement for biological tissues is extremely difficult in practical applications.The wavefront aberration of real biological tissues contains high-order aberration distortion information.These two problems can easily lead to insufficient and inaccurate training data sets.In response to this problem,we further propose a wavefront aberration detection method based on transfer learning.This method establishes the association between the target medium and the simulated medium based on the distorted point spread function(PSF)of target medium.According to the domain adaptation theory,a convolutional neural network with a two-stream structure with shared parameters is constructed.The labeled simulated medium imaging data and the unlabeled target medium imaging data are trained at the same time,and the loss function is calculated by the covariance characteristics of the two data domain.This method can effectively alleviate the influence of domain shift and high-order distortion in the feature space by aligning distribution features without prior knowledge of scattering characteristics such as labeled data of target medium.The experimental results show that compared with the method directly based on the convolutional neural network,the proposed method improves the reconstruction accuracy by 18.5% and the peak intensity of the PSF by25% with almost the same training time and processing time.The proposed method has obvious advantages in the face of thicker biological tissues and more complex wavefront aberration distortion prediction,and further expands the application scenarios of machine learning based wavefront aberration high-speed prediction.In this paper,a variety of fast wavefront aberration detection and compensation methods based on machine learning are established,which is expected to further expand the application scenarios of adaptive optics technology.It is expected to provide technical support for the construction of novel instruments for deep penetration microscopy and precise optogenetic light control with independent intellectual property rights in China.