A Sensorless Adaptive Optics Control System for Microscopy Based on Extreme Learning Machine

Imaging in vivo is of great significance in the field of biological research. Since the 21st century, adaptive optics (AO) technology has made great progress in improving the imaging quality of biological fluorescence microscopy. AO system analyzes the aberrations and outputting control parameters to the controller and then controls the modulator to compensate distortion. Nevertheless, the complicated and time-consuming conventional systems will not be applicable to deep imaging in vivo. With the development of machine learning, the convolutional neural network (CNN) has been introduced to build a rapid parametric prediction model. However, due to the nonnegligible distribution discrepancy of biological tissue, it is impossible to generalize the CNN model, which must be trained separately for different samples. To circumvent heavy calculation and long training time of CNN, in this paper, a wavefront reconstruction and correction control module based on the Extreme Learning Machine (ELM) is designed for image analysis and calculate control parameters for the spatial light modulators controller. To further demonstrate the effectiveness of our method, we compare the time consumption of model training and prediction accuracy of the trained model between ELM and CNN. The experimental results show that our proposed method improves the training speed by 13.4 times compared with CNN and achieves 87% accuracy which is the same level as CNN. The proposed sensorless AO system is of great significance for real-time in-vivo microscopy imaging.