Accurate Prediction Model Based On Improved Equivalent Uniform Dose For Radiation-induced Brain Injury In Nasopharyngeal Carcinoma

Part I:Application of Improved Equivalent Uniform Dose in Normal Tissue Complication Probability Model during RadiotherapyObjective:To investigate the effects of dose distribution and volume size on the parameter and prediction ability of equivalent uniform dose（EUD）model,and to study the application of improved EUD in the normal tissue complication probability（NTCP）model during radiotherapy.Methods:A total of 103 nasopharyngeal carcinoma（NPC）patients treated with intensity modulated radiotherapy（IMRT）and concurrent platinum-based chemotherapy were retrospectively analyzed.All patients were followed up for at least 5 years and assessed by magnetic resonance imaging（MRI）for radiation-induced brain injury.The brain was contoured primarily by automatic segmentation,and errors were corrected by manual contouring.The whole brain was divided into left half-brain and right half-brain according to the brain midline on coronal image to create a substructure with a dose distribution similar to that of whole brain but different in size.Dose volume histogram（DVH）was exported from the radiation treatment planning system by a script.By using a function,dose/volume parameters were extracted in batches from the DVHs,including dose delivered to volume V（D_V）and volume covered by doses≥D（V_D）.According to the formula EUD=（∑v_iD_i^a）^1/a,EUDs of brain and half-brain were calculated.The improved EUD was defined as the EUD calculated with V_D（representing different dose distributions）as the reference volume(EUD_VD).The independent sample t-test was used to analyze the difference in relative V_Dbetween brain and half-brain.Receiver operating characteristic curve（ROC）and area under curve（AUC）were used to evaluate the predictive ability of EUD and EUD_VDfor radiation-induced brain injury.EUD was calculated within the range of a value from 1 to 30.When the AUC was the maximal（mAUC）,the corresponding a value was selected as the optimal volume-effect parameter a.Pearson correlation analysis and linear regression analysis were used to analyze the relationship between mAUC,a and D to evaluate the impact of dose distribution on EUD,and paired sample t-test was used to compare the mAUC and a value between brain and half-brain to evaluate the impact of volume size on EUD.The D_V and V_Dparameters with the maximal AUC were selected for comparison with the improved EUD.Results:The average-DVH curves of whole brain,left half-brain and right half-brain overlapped almost exactly.The independent sample t-test showed that at the dose points of 10Gy,20Gy,30Gy,40Gy,50Gy,60Gy and 70Gy,the relative V_Ddifferences between brain and half-brain were all less than 0.01%（all P values were greater than 0.5）.As the specific dose D in EUD_VDincreased from 0Gy to 70Gy,the optimal a value of brain/half-brain decreased gradually from 21.5/22,and is always 1 when D≥55Gy.The mAUC of brain/half-brain gradually increased from 0.819/0.821,and reached the maximum when D was about 50-55Gy,and then decreased.The mAUC/a of EUD_VD（D=0-70Gy,per5Gy）in brain group and half-brain group were compared by paired sample t-test,and the difference between the two groups was not statistically significant（P=0.869/0.834）.When D≤55Gy,Pearson correlation analysis showed that D was positively correlated with mAUC（the correlation coefficients in brain group and half-brain group were 0.932 and 0.912,respectively）and negatively correlated with optimal a（the correlation coefficients in brain group and half-brain group were-0.813 and-0.877,respectively）.When D（D≤55Gy）and mAUC were included in the linear regression analysis,the R²in brain group and half-brain group were 0.868 and 0.832,respectively（P<0.001）.Within the a range of 1 to 22,AUC of traditional EUD was lower than that of EUD_V55Gyin both brain group and half-brain group（paired sample t-test,P<0.001）.Median dose of V_D(D_VD/2)was introduced as a simple substitute for EUD_VD（a=1）.When EUD_V55Gyand D_V55Gy/2were included in the linear regression analysis,the R²in brain group and the half-brain were 0.645 and 0.943,respectively（P<0.001）.In brain group,the dose-volume points with maximal AUC were D_2.5ccand V_70Gy,with AUC of 0.801 and 0.823,respectively,while the AUC of EUD_V55Gyand D_V55Gy/2were 0.853 and 0.827,respectively.In half-brain group,the dose-volume points with maximal AUC were D_1ccand V_69Gy,with AUC of0.818 and 0.827,respectively,while the AUC of EUD_V55Gyand D_V55Gy/2were0.845 and 0.844,respectively.Conclusions:Volume-effect parameter of EUD is variable and depend on dose distribution rather than volume size.Before using an empirical parameter,dose distributions of two studies should be similar.If the proportion of low dose volume is large,the predictive ability of EUD may not be better than that of simple dose/volume point.The improved EUD based on the critical dose volume improves the predictive ability and has a stable volume-effect parameter.Mean dose may be the case in which improved EUD has the best predictive ability,while the median dose may be a convenient alternative.Part II: Application of Half-brain Delineation in Prediction Model for Radiation-induced Brain Injury in Nasopharyngeal CarcinomaObjective: To study the application of half-brain delineation in prediction model for radiation-induced brain injury in nasopharyngeal carcinoma（NPC）,compare the dosimetric characteristics and predictive ability of half-brain target volume and the traditional temporal lobe target volume,and investigate the application of improved EUD based on equivalent biological dose in the prediction model of radiation brain injury under different fractional radiotherapy.Methods: A total of 220 nasopharyngeal carcinoma（NPC）patients treated with30-33 fraction IMRT and concurrent platinum-based chemotherapy were retrospectively analyzed.All patients were followed up for at least 5 years and assessed by magnetic resonance imaging（MRI）for radiation-induced brain injury.The brain and half-brain were contoured primarily by automatic segmentation,and errors were corrected by manual contouring.Temporal lobe target volumes that had been reviewed by radiotherapy specialists in the original radiotherapy plans were selected for comparison.After the DVHs of the target volumes were exported,the dosimetric parameters including D_V and V_D were extracted.Doses with different fractions were converted into equivalent dose in2Gy（EQD₂）by linear quadratic（LQ）model.EQD₂=50Gy was selected as the critical dose for half-brain,and EUDV50 Gy and DV50Gy/2 of half-brain were calculated.Coefficient of variation（CV）was used to evaluate the volume variation degrees of brains,half-brains and temporal lobes.The differences of D_V between half-brains and temporal lobes were compared by paired sample t-test.The AUC of ROC was used to evaluate the predictive power of dosimetric parameters.Logistic regression analysis was used to calculate the correlation coefficients b0 and b1 of the dose response curve for radiation brain injury.The sigmoid dose response（SDR）curve was established by Sigmoid function:P（X）=1/（1+exp（-b0-b1X））,where X was the corresponding dosimetric parameter.The tolerance dose for 5% probability in 5 years（TD5/5）and tolerance dose for50% probability in 5 years（TD50/5）were calculated.Results: The volume CVs of brains,left half-brains,right half-brains,left temporal lobes and right temporal lobes were 9.72%,9.96%,9.77%,27.85% and28.34%,respectively.Paired sample t-test showed that temporal lobe D_V was significantly lower than half-brain D_V in the range from D_max to D20cc（P<0.001）,with a reduction from 3.10% to 45.98% as the reference volume increased.The AUCs of both D_V and V_D showed an "increase-maximum-decrease" trend with the increase of volume or dose.The maximal AUCs of half-brain D_V and temporal lobe D_V were 0.828 and 0.806,respectively,and the corresponding D_V were D_1.2ccand D_0.6cc,respectively.The maximal AUCs of half-brain V_D and temporal lobe V_D were 0.834 and 0.814,respectively,and the corresponding V_D were V_{72 Gy} and V_{70 Gy},respectively.Among the four indicators,half-brain V72 Gy had the largest Youden index（0.568）with a cutoff of 0.60 cc.Further analysis of all dose/volume points showed that V_70Gy=0.86 cc,V_71Gy=0.72 cc and V_73Gy=0.45 cc cutoff points in half-brain had the same or better Youden indices（0.568,0.568 and 0.573）.The AUC of half-brain EUD_V50Gy（a=1）was 0.853,the cutoff was 59.26 Gy,and the Youden index was 0.589.The AUC of half-brain D_V50Gy/2 was 0.839,the cutoff was 58.37 Gy,and the Youden index was 0.563.According to logistic regression analysis,the predicted curves of EUD_{V50 Gy} and D_V50Gy/2were P（X）=1/（1+exp（29.825-0.476X））and P（X）=1/（1+exp（25.897-0.418X））,respectively,and the accuracy rates of prediction models were82.7% and 81.1%,respectively.The TD5/5 and TD50/5 of EUD_{V50 Gy} were56.47 Gy and 62.66 Gy respectively,and the TD5/5 and TD50/5 of D_V50Gy/2 were54.91 Gy and 61.95 Gy respectively.For comparison and reference,TD5/5 and TD50/5 of D_1.2cc were 58.7Gy and 80.0Gy,respectively.When V_{72 Gy} was greater than 0cc,the incidence of brain injury was more than 13%,and when V_{72 G} was equal to 7.66 cc,the incidence of brain injury was 50%.Conclusion: Half-brain delineation is a convenient,stable and repeatable method,which could be achieved by segmentation,and could reduce the variation of volume and improve the consistency of target volume.In a small volume range,the corresponding dosimetric parameters of half-brain and temporal lobe have little difference,so the half-brain could be used instead of the temporal lobe to predict the occurrence of radiation-induced brain injury in NPC patients.Small volume dose hotspots,such as D_max and D_{0.03 cc},may not be suitable for predicting radiation-induced brain injury.A more reasonable dose volume point is around 1cc,but may vary according to the way the target volume is delineated.The improved EUD prediction model based on half-brain can be used in different fractional radiotherapy,with larger reference volume and more dosimetric information,which can further improve the prediction ability of radiation-induced brain injury in NPC patients.