Primary Research On Classification Of Schizophrenia Based On Radiomics

Schizophrenia(SZ)is a heterogenous and complex psychiatric disorder.Patients often suffer from symptoms such as hallucinations,delusions,social withdrawal,disorganized speech,and etc.Up to now,clinical diagnosis of schizophrenia are mainly based on the subjective inquiry from clinicians with clinical questionnaires;whereas,an objective and quantifiable standard of diagnosis is still lacking.Previous studies using Magnetic Resonance Imaging(MRI)have found imaging differences between schizophrenic patients and healthy controls in different modes.What’s more,the development of radiomics has a good application in cancer and tumor diagnosis.However,it is not clear whether radiomics can be applied in mental diseases,especially schizophrenia.Inspired by this,the focus of current research work is to identify specific biomarkers that can distinguish schizophrenia patients from healthy controls,and to explore whether these brain biomarkers are predictive in combination with radiomics.The research of this paper is expected to provide an objective and quantifiable diagnostic standard for schizophrenia,which has significant and critical clinical value.Specifically,this paper consists of following aspects:First,this thesis starts with the study of imaging differences between schizophrenia patients and healthy controls.SZ patients were divided into auditory verbal hallucinations(AVHs)and non-auditory verbal hallucinations(NAVHs).Using masked independent component analysis(MICA),the independent component analysis of specific areas of interest,hippocampus can be divided into 10 sub-regions.With these 10 sub-regions as the region of interest(ROI)as well as the whole hippocampus as ROI,we analyzed the differences of hippocampal connectivity among 57 patients with AVHs,83 patients without AVHs and 71 healthy controls.Taking whole hippocampus as ROI,we found significant differences between right hippocampus and the left superior temporal gyrus as well as between right hippocampus and the bilateral superior temporal gyrus.Taking hippocampus sub-regions as ROI and comparing AVHs patients with NAVHs patients and healthy controls,we found significant differences in 8 sub-regions.In addition,correlation analysis of difference region functional connectivity values and the clinical scale showed that the resting state functional connectivity activity was significantly correlated to the clinical symptom scores of AVHs patients.These results showed that in schizophrenic patients with AVHs,subdivision of hippocampus into different sub-regions could reveal more SZ imaging differences.Second,based on the results of MICA,we further explored the classification of SZ and normal control based on characteristics of functional connectivity in order to figure outwhether the resting state functional connectivity is predictive.The data sets used in this study was collected by two different MRI machines.There were 108 first-episode SZ patients and 121 healthy controls.Among them,80% of the image data(n = 183)was randomly selected as the training set of machine learning and 20% of the data(n = 46)was randomly selected as the test set.We extracted their functional connection features through image data processing and then selected the functional connection features using the least absolute value algorithm.We further used machine learning algorithms such as LASSO to reduce dimensions of features,and then used support vector machine(SVM)to classify schizophrenia and healthy controls.In this part,we found that the accuracy rate of using functional connection features to classify and diagnose SZ patients in the data set was 87.09%,and that in the independent data set was 82.61%.In order to verify the reliability of classification accuracy,we randomly grouped and calculated classification parameters for 1000 times in this part and obtainedthe average accuracy of training and testing of 83.15% and 80.07%,respectively.Results from this part showed that functional connectivity could be used as an effective biomarker for the classification and diagnosis of SZ.Discovery of functional connectivity biomarkers is essential in the objective clinical diagnosis of schizophrenia.Third,as we had found that the functional connectivity was an effective imaging feature for clinical diagnosis,we further attempted to explore another potential identification method for schizophrenia markers through high-resolution MRI on cerebral cortex characteristics.For the sake of this,we calculated the cortical characteristics,including cortical thickness,surface area and gray matter volume,mean curvature,metric distortion and groove depth of schizophrenics and healthy controls.To be specific,we analyzed 52 first-episode schizophrenia patients and 66 healthy controls from two levels—imaging difference analysis and image group classification analysis The imaging difference analysis showed no significant difference between patient group and healthy control group.However,based on lasso dimension reduction,the average curvature(left cerebral island and left inferior frontal gyrus),cortex thickness(left fusiform gyrus)and measurement distortion(left wedge gyrus and right superior temporal gyrus)showed between-group differences and diagnostic probability.Moreover,the area under the characteristic curve was 0.88;the sensitivity,specificity and accuracy were 94%,82% and 88%,respectively.Similar results were also observed in the independent verification set,which could verify the accuracy of classification.Correlation analysis of characteristic value after dimension reduction by Lasso and clinical indexes,found that the index score obtained from multi-dimensional model was positively correlated with the severity of patients’ symptoms.i To summarize this part,we found that the imaging characteristics of cerebral cortex could also be used as biological markers for preliminary diagnosis of SZ.In addition,on the basis of above research,our further work was to extract the radiomics features of thalamus,which include the first-order features(statistical features),the secondorder features(shape texture features)and the high-order features(wavelet features).The radiomics features were applied to 191 schizophrenia patients and 199 healthy controls for classification diagnosis and clinical drug efficacy prediction analysis.In this study,we found that the classifier based on thalamic imaging features could effectively distinguish SZ patients from healthy controls,and the accuracy of classification was 68%.What’s more,we also conducted predictive analysis through radiomics features on the treatment response of SZ patients,and the degree of accuracy was 75%,which further confirmed radiomics features as effective preliminary diagnosis of SZ.This part of the results proved that the imaging characteristics of thalamus can be used to diagnose schizophrenia and to predict early treatment response of SZ patients,which has clinical significance and accuracy.In the last part,we integrated the characteristics of resting state functional connection and high-resolution structural state cortex in medical images to predict treatment effects of clinical medication used in SZ patients.To do this,we included 148 SZ patients from two different MRI data sets at baseline(85 from dataset one and 63 from dataset two).After clinical treatment,SZ patients were further divided into responders and nonresponders.In the next step,we extracted cortical features and resting state functional connection features of participants under high-resolution mode.Combining with radiomics analysis,we explored predictive performance of the mixed features for the clinical treatment of SZ.Following this,we analyzed prediction models based on structural features,functional features and combined features.Results showed that,the prediction accuracy of this model was 80.38% when using the resting state function connection feature alone,69.68% when using the highresolution cortex feature alone,and 92.04% when combining the two modality image group features.This part of the results confirmed that high-resolution structure image and resting function image,as the prediction model of the early treatment response of SZ,can optimize the effective treatment strategy.In conclusion,this paper applies radiomics strategy and technology to the preliminary diagnosis and early treatment of schizophrenia.In addition,this paper provides a more objective biological index for clinical diagnosis of schizophrenia by using the method of image analysis combined with multi-modal MRI.The research design of this paper are as follows: 1)to find out specific imaging features of SZ and health control,2)to extract imaging histograms under different modes specific to SZ patients,and use machine learning method to carry out preliminary diagnosis of SZ patients,3)to conduct preliminary therapeutic prediction analysis for SZ patients based on the imaging histology analysis method.