Research On Centerline Extraction In Neuron Images Based On Flux Model And Deep Neural Networks

Brain neuronal morphology contains a wealth of information.Reconstructing neuronal morphology from optical microscope image stacks is of great value in understanding brain-like mechanisms and functions,analyzing disease pathology,mapping brain atlas,etc.Among them,neuron centerline extraction provides an effective morphological representation of the axonal and dendritic structures of nerve fibers,which is a critical step in neuronal reconstruction.The accuracy of centerline extraction will directly affect the subsequent reconstruction steps,such as topological connection and radius estimation.However,the current neuronal centerline extraction still faces challenges,such as weak neuronal image signals,high noise levels,and complex neuronal fiber structures.Therefore,this paper proposes a brain neuron centerline extraction algorithm based on the flux model and deep networks.The main research is as follows.Firstly,this paper proposes a 3D Tubular Flux Model to represent the structural information of neurons and generate robust flux features.Using the neuron centerlines manually annotated by experts,the model first approximates neuronal fibers as 3D tubular regions,and then encodes the voxels within the centerline and its neighborhood as vectors.Through quantitative analysis of the vector field of the three-dimensional tubular region,the distance distribution and direction distribution are calculated,and the flux characteristics are finally obtained.Therefore,the 3D tubular flux model proposed in this paper fully mines the internal structural information of neurons,and converts a single centerline position into a flux feature that reflects the contextual relationship.And the flux feature can be used as a supervisory signal in various neuron centerline extraction tasks based on machine learning algorithms.Secondly,this paper designs a neuron centerline extraction algorithm based on a two-stage deep network,which is suitable for more unlabeled scene neuron centerline extraction tasks.In the first stage,flux features serve as supervisory signals to guide a 3D Convolutional Neuronal Network(3D CNN)to predict flux features from neuron image stacks.In the second stage,neuron centerlines are extracted from the flux features predicted in the first stage using a lightweight U-Net.Finally,a spatially weighted average strategy is used to optimize the network output to suppress the centerline multi-voxel width response.Therefore,the algorithm uses flux features as intermediate supervision information to guide network learning,which can avoid extracting redundant features and achieve more accurate neuron centerline extraction.Finally,experiments are conducted on two challenging neuron datasets to verify the effectiveness of this algorithm.Compared with the state-of-the-art centerline extraction algorithms,the proposed algorithm achieves the best performance currently and can extract the complete neuron structure under the interference of image noise.At the same time,combining this algorithm with the existing neuron reconstruction algorithm effectively improves the result of neuron morphological reconstruction.