Tree-Like Structure Junction Detection In Biomedical Images Based On Improved U-Net

The junction points of biomedical images play an important role in biomedical research,such as retinal biometric recognition,retinal image registration,eye-related disease diagnosis and brain neuron reconstruction.However,it is extremely challenging to automatically detect the junctions in original biomedical images without segmentation,because the branch structure of biomedical images(such as retinal and neuron images)is usually complicated and the contrast between foreground and background is very low.Aiming at the difficulty of junction detection in original biomedical images which contain weak filamentous signals,a dual-channel O-shape neural network architecture with Attention modules,named Attention O-Net,is proposed to detect junctions in biomedical images without segmentation.This architecture consists of two channels: Junction Detection Channel(JDC)and Local Enhancement Channel(LEC).The main research contents are as follows:Firstly,in the junction detection channel,the heatmap indicating the probabilities of junctions is estimated via improved U-Net and followed by choosing the positions with the local highest value as the junctions.However,it is challenging to detect junctions when the images contain weak filament signals with low contrast.Therefore,in this paper,a local enhancement channel is constructed to enhance the thin branch foreground and make the network pay more attention to the regions with low contrast,so as to better detect the thin junctions.At the same time,a radius adaptive label is designed to train the local enhancement channel.This label expands the proportion of thin branch foreground and reduces the number of thick branch foregrounds,which helps to alleviate the imbalance between the proportion of thin branch and thick branch foreground,accelerate the speed of network convergence and reduce the difficulty of network training.Secondly,to make full use of the information of the local enhancement channel and the radius adaptive label,the attention modules are used to introduce the decoder feature map of the local enhancement channel into the junction detection channel,so that the two channels can interact many times.Because the local enhancement channel is trained with the radius adaptive labels,this channel carries rich and clear thin branch structure information and boundary information,which can correct the error of setting the background as the training target in the heat map label in the junction detection channel.Through the attention module,the feature map of the local enhancement channel is introduced into the junction detection channel,so as to use the unique element-wise multiplication in the attention module to reduc e network noise interference,establish a complementary relationship between the two channels,further integrate local features and context information,and take full advantage of the local enhancement channel.Finally,experiments on two retinal data sets and one neuron data set show that can accurately detect the weak signal junctions of the tree structure in biomedical images.Compared with all other comparison methods,Attention O-Net achieves the highest F1-scores.In addition,experiments also show that the junctions detected by Attention O-Net are helpful to promote the development of retinal biometric recognition technology and neuron reconstruction algorithm.