Research On Biomedical Image Analysis Based On Time Series Deep Learning Ray Segmentation Algorithm

With the development of imaging technology,the biomedical images obtained by various imaging devices play an important role in biomedical research and clinical medical research.Automatic analysis of biomedical images using computers,such as biomedical image segmentation and biomedical image shape analysis,is increasingly important in understanding normal functions and pathological processes in biology.However,most of the existing biomedical image analysis algorithms are proposed for a specific image analysis task.In this thesis,based on the time series deep learning ray segmentation algorithm,we design an algorithm that can serve multiple image analysis tasks.First,since biomedical images often contain strong background noise and weak-signal structures,this paper proposes a biomedical image segmentation algorithm based on the ray-shooting model and the Dual-Channel Bidirectional Long Short-Term Memory(DC-BLSTM)network to enhance the weak-signal structures and remove background noise.Specifically,the ray-shooting model is used to extract the voxel-intensity features and boundary-response features within a local region of the image.And the thesis designs a neural network based on the DC-BLSTM to detect the foreground voxels according to the intensity distribution features extracted by multiple ray-shooting models that are generated in the whole image.This way,we transform the image segmentation task into multiple 1D ray/sequence segmentation tasks,which makes it much easier to label the training samples than many existing Convolutional Neural Network(CNN)based biomedical image segmentation methods.Since the designed features take neighbor information into consideration,and the DC-BLSTM is able to summarize the feature distributions along a ray,it ensures the robustness of ray segmentation.Then,since other biomedical image shape analysis tasks,such as vessel width estimation and 3D neuron shape reconstruction,need to cast longer rays and more accurate ray segmentation/termination,the Deep Rayburst algorithm is proposed based on the above ray segmentation method.Specifically,to cast longer rays to ensure that all rays can reach the surface of tree-like structures,the Dual Channel Temporal Convolutional Network(DC-TCN)is used to terminate the rays instead of the DC-BLSTM,because the TCN module is capable of maintaining a much longer effective memory than LSTM with the same capacity.Furthermore,to obtain more robust ray termination results,a Multi-Feature Rayburst Sampling(MFRS)is designed by extending each ray of the Rayburst to multiple parallel rays which extract a set of feature sequences.A Gaussian kernel is then used to fuse these feature sequences and outputs a fused feature sequence with more contextual information.Based on the extracted fusion feature sequences,DC-TCN can adaptively terminate the rays on the surface of tree-like structures.By analyzing the distribution patterns of the terminated rays,the algorithm can serve multiple shape analysis tasks including soma shape reconstruction,neuronal shape reconstruction,and vessel caliber estimation.Finally,experimental results on multiple biomedical image datasets demonstrate that the proposed ray segmentation-based biomedical image analysis algorithm outperforms existing state-of-the-art methods,and can effectively promote the research on automatic biomedical image analysis.