Applying Machine Learning Methods To Whole Brain White Matter In Patients With Nasopharyngeal Carcinoma After Radiotherapy

Objective:Using magnetic resonance diffusion tensor imaging（DTI）,based on the fiber bundle of spatial statistics（TBSS）and machine learning methods in different periods after radiotherapy in patients with nasopharyngeal carcinoma（NPC）whole brain white matter structure and whole brain white matter changes in network circumstance,exploring based on whole brain DTI data,can machine learning methods in different periods after radiotherapy patients to distinguish with the control group at the individual level,to achieve high recognition rate,finding the most distinguish ability of cerebral white matter and white matter connections,as biological marker of radiation brain injury,provide theoretical basis for clinical diagnosis and early intervention.Materials and methods:DTI scan was performed in 77 patients after radiotherapy with nasopharyngeal carcinoma and in the normal control group（n=67）.1.DTI data preprocessing:transform the original DICOM format data into NIFTI format,using McFliirt to eliminate head error,distortion correction,using BET for brain extraction,and then generate the FA diagram by DTIFit.2 image processing:then use TBSS to find the difference between the FA values of white matter in the case group after radiotherapy of nasopharyngeal carcinoma and control group.The obtained FA values are matched with each other to get the target image,and the image is registered into the MNI.152 space to create an average FA value skeleton.3 Classification:connect the FA skeleton image into the feature vector,extracted significant different characteristics between the groups,and use support vector machine（SVM）to classify.4 white matter network construction:use the automatic anatomical marker template（AAL）to carry out brain segmentation on the DTI image of each participant（116 cerebellum）,and conduct deterministic tracking of white matter fibers,then generate a symmetrical 116*116 brain matrix,remove the diagonal components,choose the triangle selection component（6670 elements）as classification features.5 white matter network classification:use the two sample t test to extract the significant different characteristics between the groups,and carry out extraction of nonlinear features,and finally use SVM classification.The reliability of the classifier was evaluated by permutation test and ROC curve.Results:0 to 6 months after radiation therapy group and control group classification recognition rate is 84.5%,6 to 12 months after radiation therapy group and control group classification recognition rate is 83.9%.>12 months after radiation therapy group and control group classification recognition rate is 74.5%Compared with control group,0 to 6 months after radiotherapy group has the ability to distinguish between brain areas,of which the FA values are reduced,mainly in bilateral cerebellum,including Cerebelum 7b_L and Cerebelum_Crusl R.Compared with control group,6 and 12 months after radiotherapy ’group has the ability to distinguish between brain areas,of which the FA values are reduced,the main area is located in the left temporal lobe white matter and the left cerebellum,including Cerebelum_Crusl_L,Cerebelum_8L,Temporal_Mid_L.The results show that compared with the control group,.>12 months after radiotherapy group has the ability to distinguish between brain areas,of which the FA values are reduced,including Cerebelum_Crusl_L,Cerebelum_8L,Temporal_Mid_L and Temporal_Pole_Sup R.For whole brain white matter connections,0 to 6 months after radiotherapy group and the control group:the SVM classifier classification recognition rate reached 82.5%（SS=83.3%,SC=83.3%;p<0.0001）;6 and 12 months after radiotherapy group and the control group:the SVM classifier classification recognition rate reached 78.4%（SS=76.7%,SC=76.7%;p<0.0001）;>12 months after radiotherapy group and the control group:the SVM classifier classification recognition rate reached 76.3%（SS=80%,SC=80%;p<0.0001）.Compared with control group,each group after radiotherapy of brain white matter connections（consistency）are reduced,all three groups does not match,but are mainly distributed in the frontal lobe-edge network,temporal lobe and cerebellum.Conclusions:This study assessed using DTI-TBSS and machine learning methods nasopharyngeal carcinoma after radiotherapy in patients with brain white matter microstructure and whole brain white matter network changes,the results found that group after radiotherapy and control group both can be distinguished between each other,and can achieve high recognition rate.Radiation brain injury is a disease of the whole brain white matter network anomalies.The strength of most consistent ability（white matter connections）are reduced,distributed mainly in the frontal lobe-edge network,temporal lobe and cerebellum.These brain white matter and brain white matter connection mode or can be used as a biological marker for clinical early diagnosis and treatment of radiation brain injury.