Research On Automatic Focusing Algorithm For Single-Cell Mass Spectrometry System Via Deep Learning

Single-cell mass spectrometry(MS)technology,as a non-targeted detection technique,has excellent characteristics such as high sensitivity and high specificity.It is widely used in cellular metabolomics analysis.Therefore,in-depth research on single-cell MS technology is of great value and significance.However,the automatic positioning of the pipette in the single-cell MS system,specifically the automatic focusing problem in the microscopy imaging module of the system,is the most challenging issue.Existing automatic focusing algorithms cannot achieve precise positioning of the pipette tip area,and they also fail to meet the real-time requirements of the single-cell MS system.Therefore,a precise and efficient automatic focusing algorithm would be beneficial for obtaining more accurate MS analysis results and improving the overall operational efficiency of the single-cell MS system.In this study,we addressed the problems in the automatic focusing algorithm of the single-cell MS system by utilizing deep learning techniques.The specific research contents include:(1)A focus position prediction model based on ROI features and temporal networks.The collected image data is preprocessed,including cleaning,labeling,and data augmentation.The Faster R-CNN(Region-based Convolutional Neural Networks)object detection network is used to locate the target regions in the images and extract their features.This approach solves the focusing interference caused by image noise and a large amount of background information.Based on the characteristics of the data and considering the practical application scenario of the focusing algorithm,the features of the target regions in three consecutive images are used as inputs to the temporal neural network.The temporal neural network extracts information on the changes in the degree of focus from the features and predicts the motion state of the robotic arm at the next moment.This allows the pipette to gradually reach the focusing position,ultimately achieving automatic focusing in microscopy imaging.Compared with existing automatic focusing algorithms in single-cell MS systems,the proposed algorithm reduces the operating time by 381 ms and improves the focusing accuracy within the depth of field by 2.5%.(2)A defocus depth estimation model based on ROI feature distillation and label distribution learning.The traditional focusing depth method requires computing each image during the focusing process,which undoubtedly increases the time consumption and decreases the overall operational efficiency of the single-cell MS system.To address the issues of time consumption and the influence of high-resolution images on algorithm speed,a lightweight regression model based on label distribution learning is used to estimate the defocus depth.The model takes a single image as input,and to accelerate the inference process,the image is processed in blocks after removing the background.The distribution loss and weighted regression loss are utilized to train the regression model.To further improve the accuracy of the regression model,a feature distillation mechanism is introduced.The target region feature extraction network is used as the teacher model,while the regression model serves as the student model.After distillation,the corresponding focusing error of the model is reduced by 0.191,and the focusing accuracy within the depth of field is improved by 6.68%.To compensate for the accuracy issue,a two-step focusing strategy is employed.Compared to the focusing position prediction model,the algorithm achieves a significant improvement in focusing efficiency,while maintaining a focusing accuracy of 96.94% within the depth of field.