Medical Diagnostic Technology Development Of Spectral Signal Combined With Multi-fusion Strategy

Due to its advantages of less trauma and higher sensitivity,biofluid spectroscopy has been widely applied in the field of medical diagnosis.However,there are still some problems in the spectral signal processing,such as not designing the data processing algorithm with higher matching degree for the spectral data.The types of research objects are single,and the classification models for diseases and healthy people cannot cope with the complexity of the environment and the complexity of the research objects in the clinical diagnosis.In addition,data comes from a single source.Many researchers at home and abroad build models based on a single spectral data,which greatly limits the robustness of models in complex environments and makes it difficult to build multiple classification models due to a single feature.These problems make the technique of spectroscopic diagnosis still far from clinical application.Based on these problems,in this paper,I carries out four parts of research from the perspective of data fusion:1.The diagnostic model was constructed by combining the multi-humoral Raman spectra of serum and urine with various machine learning algorithms such as support vector machine(Supported Vector Machine,SVM)for chronic renal failure,with the highest accuracy of 89.7%.In addition,we further explored the potential of deep learning networks combined with Raman spectroscopy to construct classification models.Based on thyroid disease data,the classification model of thyroid dysfunction/control samples and the classification model of hyperthyroidism/hypothyroidism/control samples were established.Among the classification model of thyroid dysfunction/control samples,the accuracy of the model constructed by principal component analysis(Principal Component Analysis,PCA)combined with multi-feature fusion convolutional neural network(Multifeature Fusion Convolutional Neural Networks,MCNN)reached 94.01%.In the model of hyperthyroidism/hypothyroidism/control samples,we adjusted Alex Net to be more suitable for processing spectral data,and the final accuracy was as high as 89%.In this study,the model was optimized to promote the application of Raman spectroscopy combined with artificial intelligence algorithm in the diagnosis of hyperthyroidism and hypothyroidism.2.At present,it is difficult to carry out multi-disease diagnosis technology based on the single data source of spectral analysis.Based on this situation,the fusion diagnosis technology based on multi-spectroscopy is proposed in combination with the complementarity of infrared spectroscopy and Raman spectroscopy.Spectral fusion diagnosis based on thyroid function of study,we have different spectra to extract the features of fusion to build classification models such as extreme learning machine(Extreme Learning Machine,ELM).The final results of abnormal thyroid function show that the spectral fusion diagnosis model relative to the accuracy of the single spectrum has certain promotion;In addition,we have further developed the research of spectral fusion technology in the multi-disease diagnosis of lung cancer,glial cancer and esophageal cancer.The four classification model and the dicclassification model were constructed respectively.For different classification models,two methods of low-level fusion and feature fusion are constructed.The classification algorithms include: SVM,CNN(Convolutional Neural Networks,MCNN)and CNN-LSTM.In the fourclassification model,the accuracy of the model after feature fusion is 90.23%,88.37%and 83.72%,respectively,which is 11.63%,10.23% and 3.25% higher than that of the single-spectral four-classification model.The effectiveness of the low-level fusion model has also been significantly improved.In the dichotomy,the results of feature fusion and low level fusion are consistent with that of the quaternary classification.3.In this study,a multi-classification model was established for cervical inflammation,low-grade squamous intraepithelial lesion(Low-Grade Squamous Intraepithelial Lesion,LSIL),high-grade squamous intraepithelial lesion(High-Grade Squamous Intraepithelial Lesion,HSIL),cervical squamous cell carcinoma and adenocarcinoma based on the features of tissue Raman spectrum fusion Fourier transform.Compared with the single spectral model,the overall performance of the Fourier-Raman spectral feature fusion model constructed based on CNN-LSTM and other algorithms has been steadily improved,which provides a new method for the future medical aided diagnosis and other fields to further explore spectral information and improve the classification accuracy of the model.4.Based on the instability of the current spectral deep learning model and the lack of data,this study constructed a transfer learning source network by selecting representative diseases such as thyroid dysfunction and hepatitis B.With the spectral data of hepatitis C as the target data set,the single transfer learning model was constructed respectively.On this basis,we used the output probabilities of different single transfer learning models as the input of logistic regression to construct the decision level fusion model,and the results show that the classification accuracy has been further improved.In the end,the experiment proved that the unrelated Raman data sets with different scales were still intrinsically connected,which also laid a foundation for us to construct a spectral multi-transfer learning fusion model with high stability and strong data inclusion.