Neuron Similarity Measurement Method Guided by Local Feature and Point Cloud Registration

Aiming at the problem that the existing neuron morphological matching and similarity measurement methods are difficult to effectively handle large-scale and complex neuron data, a neuron similarity measurement method is proposed based on local features and point cloud registration. Firstly, global features are used for large-scale retrieval to complete the initial selection of neuron similar data. Secondly, unsupervised local features are extracted according to the selected neuron data, based on the deep convolutional autoencoder, in achieving the coarse registration between pair-wised neurons. Subsequently, the matching problem of neurons with spatial tree structure can be transformed into 3D point cloud registration problems based on iterative closest point algorithms. Finally, the similarity measurement of neuron morphological data can be achieved through the fusion between the registration score and the retrieval results under global features. 19 286 neurons are extracted from the NeuroMorpho public data set and compared with 6 other similarity measurement methods. The accuracy of Top-1 and Top-50 of the proposed method are 0.981 and 0.721 respectively, both of which are better than the existing methods. Experimental results on large-scale datasets demonstrate the effectiveness and accuracy of the proposed method.