Cryo-electron microscopy(Cryo-EM)is becoming more and more widely used in structural biology.In the process of 3D reconstruction of molecular structure by cryo-EM,the selection quality of particles in micrographs affects the quality of subsequent particle classification and analysis,which is directly related to the resolution of the final 3D reconstruction.Deep learning algorithms in general object detection are studied including the structure of convolutional network,the classification algorithm and the bounding box regression mechanism of object detection.On this basis,a convolutional neural network method called RSelector is proposed for particle selection tasks in cryo-EM micrographs.Analysis of the bounding box regression optimization process in general object detection shows that the loss function calculation method directly predicting the scale information will inevitably lead to the differentiation of the prediction performance of different scale objects.On this basis,the scale-normalized coordinate regression loss function is proposed.The regression loss function improves the object bounding boxes regression algorithm in object detection,and performs the same-rate bounding box coordinate regression for objects of different scales.Experiments in several public cryo-EM microscopy data sets show that the RSelector method provides better particle selection performance than current several common methods,and scale-normalized regression loss function improves the accuracy of predicted position in particle selection tasks. |