Neonatal Pain Assessment Based On Video Analysis

Repeated pain stimulation(such as immunization,plantar blood collection and surgery,etc.)on the neonates can cause changes in the nervous system,which will lead to stunting and learning disabilities.Pain assessment is an important cornerstone of pain management.Therefore,it is of great significance to improve the quality and efficiency of neonatal pain assessment for the healthy growth of newborns.In the thesis,a new exploration of neonatal pain expression recognition technology is carried out,focusing on the deep three-dimensional convolutional neural network algorithm in the field of deep learning,and successfully applied to the neonatal pain assessment based on video analysis.The main work is as follows:(1)A neonatal facial expression video database is established.Collecting videos with neonatal facial expression in different states through communication and cooperation with the the children's hospital.After pre-processing operations such as editing and normalization,a neonatal facial expression video database containing four expressions of calm,crying,slight pain and severe pain has been established.(2)A method of neonatal pain assessment based on three-dimensional inflated convolutional neural network(Inflated-3D ConvNet,I3D)is proposed.I3 D expands the basic two-dimensional convolutional neural network into a three-dimensional convolutional structure.It can simultaneously extract the spatial and temporal features of neonatal expression video.In addition,I3 D adopts multiple 3D convolution kernels of different sizes to extract features of different scales and obtain better feature representation.(3)A method of neonatal pain assessment based on deep three-dimensional residual network(Deep Residual-3D ConvNet,DR3D)is proposed.The core of the algorithm is to decompose three-dimensional convolution into 2D spatial convolution and 1D temporal convolution,which greatly reduces the huge number of parameters brought by three-dimensional convolution,reduces the computational difficulty of the deep network.In the thesis,the experimental method of fine-tuning training is used,and the recognition rate on the neonatal facial expression video database is 66.54%.(4)A method of neonatal pain assessment based on three-dimensional temporal convolutional neural network(Temporal-3D ConvNet,T3D)is proposed.T3 D simulates the three-dimensional convolution kernel of variable temporal dimension through the temporal transition layer,and connects the temporal features of different temporal dimensions,which can effectively fuse the short-term,medium-term and long-term temporal information of neonatal facial expression videos.This method finally achieves a good recognition rate in the neonatal facial expression video database.