Detection Of Branch Point In Biomedical Images Based On Deep Learning

In biomedical research,the biomedical images obtained by various imaging devices can provide a more intuitive and reliable basis for biological research and clinical medical diagnosis and treatment.For tree-like structures,such as neurons,retinal vessels,and bronchi,the automatic detection of branch points is a key technology of great significance.Because branch points can not only effectively improve the morphology reconstruction,but also play an important role in cell morphological evolution and calibration,pathological analysis of related diseases and brain atlas construction.This paper first analyzes the relevant algorithm theories and experimental methods of feature points detection in natural images and biomedical images,and summarizes the detection effects and shortcoming of these methods.After conjoint analysis of the characteristics of branch points detection tasks in biomedical images and related deep learning theories,the new idea for 2D/3D branch points detection in biomedical images based on deep learning is proposed.A multi-task learning based model is proposed for branch points detection in 2D biomedical images.By regressing the branch heatmap in the main task of the multitask model,the problem that the networks are difficult to converge can be avoided,which is the main problem for using convolutional neural networks to directly regress the coordinate of branch points.At the same time,considering the distribution characteristics of branch points in biomedical images,an auxiliary segmentation task is constructed to improve the recall of branch points detection.The shared layers of the multi-task model are designed based on the U-Net,to make good use of semantic information of different levels.And the ideas of deep separable convolution and residual learning are integrated,which not only improve the performance of the networks,but also decrease the number of parameters,making the training of the networks to be easier.The test results on the retinal vessel images in the Digital Retinal Images for Vessel Extraction(DRIVE)dataset show that the proposed 2D branch points detection method can accurately detect the branch points in the biomedical image.For 3D biomedical images,the model is extended for 3D branch point detection,and the attention mechanism is incorporated.The addition of the attention mechanism enables the model to dynamically adjust the weights of different positions of the feature map,thereby suppressing unimportant information in the 3D branch points detection and reinforcing useful information.Then,the detecting results are fed into multi-scale multi-view networks(MSMV-Net)for false positive reduction.In this paper,3D branch points are labeled on the bronchi images of ANODE09 dataset and the neuron images of Bigneuron dataset respectively,and the experimental results on those images demonstrate the effectiveness of the proposed algorithm.