The Research Of Burn Depth Detection Based On Feature Learning Of Near Infrared Spectral Images

Accurate burn depth detection of patients is a very important research direction.The diagnosis results are related to the incidence of subsequent wound infection and hypertrophic scar.Currently the primary methods for burn depth assessment remain a subjective evaluation by clinicians and its accuracy of clinical diagnostic just has only 65%-70%.Therefore,there is an urgent need for a non-invasive and efficient method to detect burn depth.Near Infrared Spectral Imaging(NIRSI),as a non-invasive and non-contact spectroscopy technology,can detect the structure and composition of burn skin,so it can be used for diagnostic analysis of skin burn depth.However,this method has less research in the field of burn depth detection,which limits its application in burn depth detection.In order to solve the above problem,the thesis focuses on how to use the feature learning method combined with near-infrared spectral images to accurately predict the burn depth and proposed the method of burn depth detection based on feature learning of near infrared spectral images.The main contents of the thesis are as follows:(1)For the full-field quantitative detection of burn depth,the RSER-KNN ensemble regression model based on near-infrared spectroscopy images was proposed to realize burn depth detection.Firstly,the NIRSI instrument is used to collect the full-field spectral images of the burn wounds.After pre-processing the data,the fusion algorithm of Principal Component Analysis(PCA)and Linear Discriminant Analysis(LDA)were used to learn feature from the spectral signals,and the subspace ensemble idea was used to train multiple K-nearest neighbors(KNN),which can improve the accuracy of burn depth prediction.According to the prediction results of the ensemble regression model algorithm,the depth map of the whole field burn skin is presented to provide a visualization result.The map of the optical property parameters of burn wound was extracted by the diffuse reflection theory to analyze the changes of tissue structure components for different burn depths and carry on the mechanism study of burn depth detection based on near infrared spectral images.(2)For the burn depth detection in cross-domain samples of burn wounds,the deep transfer learning model based on near-infrared spectroscopy images was proposed.Firstly,the NIRS instrument was used to collect the spectral images of abdomen tissues from pig as the A and the B dataset,and they were respectively treated as source and target domain data to conduct experiments.The Convolutional Neural Network(CNN)regression model was pre-trained and validated on the A dataset,and compared the experimental results with traditional machine learning,which can verify the feature learning ability of CNN model and its efficient prediction ability for burn depth.Then,the deep transfer learning method was introduced to overcome the poor generalization ability of the trained model caused by the different distribution in cross-domain burn dataset.Transfer learning method would transfer the bottom layer structure and parameters of the pre-trained CNN regression model,and added the optimized top layer structure to form the transferred CNN network(CNN-transfer).Finally,a small number of samples of the B dataset were used to fine-tune the CNN-transfer to improve the depth prediction accuracy of the B dataset,realizing an accurate burn depth prediction in cross-domain samples.The thesis provided a new thought and solution for the full-field quantitative detection of burn wounds and the burn depth detection in cross-domain.It provided the theoretical basis for the burn depth detection based on near-infrared spectroscopy combined with feature learning,so it has a certain theoretical significance and application value.