A Radiomics Model From Longitudinal MRI Texture Features For The Prediction Of Radiation-induced Temporal Lobe Injury In Nasopharyngeal Carcinoma

Background:Nasopharyngeal carcinoma(NPC)is a common cancer in East Asia with a regional difference character.It occurs more frequently in Southern China,specifically in the Guangdong province.The radiation therapy serves as the most effective treatment for nasopharyngeal carcinoma patients.As the lower temporal lobe is within the RT target volume,RTLI is a major neurologic complication following RT for NPC.At present,the diagnosis of RTLI mainly depends on conventional CT and MR images,however,the diagnostic capabilities were not satisfactory,when conventional MR shows brain white matter edema and demyelinating performance indicated the disease is often at the late stage.On the other hand,patients with normal imaging may appear apparent nervous system symptoms,or show the loss of cognitive function in neuropsychological testing experiment.Recently,the term "radiomics" has attracted increasing attention,which is known as the process of the conversion of medical images into high-dimensional mineable data through high-throughput extraction of quantitative features,followed by subsequent data analysis for decision support.The core hypothesis of radiomics is that the models,which can include biological or medical data,can provide valuable diagnostic,prognostic or predictive information.Purpose:To predict impending radiation-induced temporal lobe injury(RTLI)in nasopharyngeal carcinoma(NPC)patients by evaluating the temporal lobe with muti-parametric MRI based radiomics.Methods and materials:From January 2006 to August 2016,a total of 200 consecutive nasopharyngeal carcinoma patients(143 men and 57 women;mean age,50 years ± 10.98)with RTLI confirmed on magnetic resonance imaging(MRI)in our hospital were enrolled.All patients had underwent baseline MRI and MR for follow-up.After completion of radiation therapy,all patients had underwent regular MR examination in the latent period,and they were followed up every 3-6 months during the first year,and every 6 months in the second year,and then annually thereafter until RTLI onset,including CET1-weighted(CETl-w)and T2-weighted(T2-w)image.The latency is defined as the interval from radiation until first diagnosis of RTLI and the latent period is defined as N(years),and all MR examinations are classified based on the follow-up time backward as groups of N-1,N-2,N-3 and so on,the group less than 50 times could not be included in the study.We achieved axial contrast-enhanced T1-weighted Digital Imaging and Communications in Medicine(DICOM)images and T2-weighted DICOM images in the institutional Picture Archiving and Communication System(PACS,Carestream,Canada).We used ITK-SNAP software for three-dimensional manual segmentation(open source software;www.itk-snap.org).The region of interest(ROI)range covered the middle and lower lobe of the temporal lobe,delineated from the slice of cerebral peduncle to the lowest slice of temporal lobe both on the axial CET1-w and T2-w images.All feature extraction methods were implemented using MatLab 2014a(MathWorks,Natick,MA,USA).Radiomics features include 4 types of non-texture features and 43 types of texture features.Using five fold cross validation method to determine the number of training set and validation set.The process of our method mainly consists of four steps:processing of the MRI data,feature extraction,feature selection,construction of the final prediction model for validation.The evaluation is conducted on the whole temporal lobe,grey matter and white matter of temporal lobe using CET1-w,T2-w and the combination of CET1-w and T2-w seperately.The prediction performance was then estimated based on the combination of LDA and 0.632+ bootstrap AUC.Results:Median latency was 38 months(range,9-97 months),we collected 1383 times of baseline and followed MR.105 patients had unilateral RTLI,and 95 had bilateral RTLI.There are 200,171,147,104,72,50 times of MR examinations in groups of N-1,N-2,N-3,N-4,N-5,N-6 respectively.A total of 4 non-texture and 10320 texture parameter features were extracted from each scan.Of these,we selected 2 features from CET1-w images and 14 features from T2-w images that yielded the best predictive performance.No differences were found between the training and validation cohorts in terms of age,gender,overall stage,or latency(p=0.314-0.876).As time goes from the back forward,the prediction performance showed a decrease trend,and the group of N-1 achieved the best performance.If the numbers of selected features were reached to 15,the prediction model with combination of CET1-w and T2-w for whole temporal lobe of group N-1 yielded the best AUC of 0.73 ± 0.002 in the training cohort,which was confirmed to be 0.76 ±0.001 in the validation cohort.In the validation set,the predictive performance of features derived from joint CET1-w and T2-w outperformed those from T2-w and CET1-w separately(AUC:0.76 Vs 0.73 Vs 0.65).The features of whole temporal lobe showed higher predictive performance than those of white matter or grey matter alone,either for CET1-w(AUC:0.65 Vs.0.58 Vs.0.60)or T2-w(AUC:0.73 Vs.0.61 Vs.0.66).Conclusions:Full utilization of temporal lobe textural features extracted from MRI images possess great promise for the early prediction of impending RTLI,affecting the clinical strategies in practicce.It may predict impending RTLI within next year,advising clinical docotor to shorten the time of review or treatment in advance.As the radiomics signature from joint CET1-w and T2-w images of whole temporal lobe demonstrated better prognostic performance than those from either CET1-w or T2-w images separately,it is recommended that every patient should have contrast-enhanced scan for follow-up.