Evaluation And Prediction Of Radiation-induced Brain Injury In Patients With Nasopharyngeal Carcinoma Based On Brain Network

Nasopharyngeal carcinoma(NPC)is one of the most prevalent malignancies in North Africa,Southeast Asia,and southern China.Radiotherapy(RT)is a dominant treatment for NPC patients,but it can cause damage to the adjacent brain tissue,and even lead to temporal lobe necrosis(TLN).TLN is irreversible and seriously affects brain cognition,but the early changes induced by radiation are curable.Therefore,if early brain alternations can be used to predict TLN,the risk of TLN can be reduced by intervening in early brain dysfunction.Radiation-induced brain injury(RBI)involves abnormalities in local brain regions and large-scale networks,suggesting that the brain network may be a sensitive marker of RBI.This study aims to explore radioactive functional brain network in NPC patients and construct a prediction model of TLN based on dosimetric parameters and early multimodal brain network metrics.Our work mainly included:(1)We explored the longitudinal changes of radiation-induced functional brain networks in NPC patients,and constructed a prediction model of brain dysfunction.Thirty-six NPC patients and fifteen healthy control(HC)subjects were included in this study and functional magnetic resonance imaging data were collected at baseline,at 03 months,6 months,and 12 months after RT.Global and nodal changes in the brain functional network after RT were observed by the graph theory method and mixed design analysis of variance.The results showed that the nodal efficiency(NE)in bilateral frontal lobes,temporal lobes,and right insula presented a significant "declinerecovery" pattern over time in NPC patients.In addition,a normal tissue complication probability(NTCP)model was constructed using clinical and dosimetric characteristics to predict brain dysfunction.The results showed that the NTCP model predicting NE changes in the right superior temporal gyrus performed relatively well(area under the curve,AUC=0.68),and dosimetric parameters(D20.0cc and V20.0)acted as a significant part in this NTCP model.(2)We explored the feasibility of brain network for predicting TLN in NPC patients.The study included sixty-three NPC patients with three years of follow-up period.The longitudinal diffusion and functional magnetic resonance imaging data were collected at baseline and at 0-6 months after RT(acute phase).The prediction model was constructed by support vector machine(SVM)algorithm.Clinical(demographic,clinical stage,and dosimetric parameters)and imaging characteristics(topological metrics and coupling coefficients of structural and functional networks)were used as the model features.The AUC was calculated for model evaluation,and the DeLong test was performed for model comparison.The results showed that the model combining clinical and multimodal MRI features(AUC=0.915)performed better than the model only using clinical(AUC=0.806,p=0.042)or multimodal MRI features(AUC=0.664,p=0.003).In the model combining clinical and MRI features,D0.1cc,D20.0cc,V50.0,and degree centrality of STG in the acute phase(acute_dMRI_DC_STG)were the most predictive characteristics.In addition,curve fitting was adopted to explore the relationship between the above features and the TLN risk.We obtained the dose tolerance for 50%probability of TLN development in 3 years,with 76.08 Gy for D0.1cc,31.34 Gy for D20.0cc,and 7.37%for V50.0.Furthermore,the TLN risk increased significantly when acute_dMRI_DC_STG reached the cut-off point of 20.In this study,we found that radiation-induced brain network damage begins in the acute phase and recovers over time.A prediction model of brain dysfunction based on dosimetric and clinical parameters was proposed.In addition,we found that dosimetric parameters and topological metrics of brain networks are crucial biomarkers of TLN.The machine learning model combining clinical and multimodal MRI network features can accurately predict TLN,which has important significance for early diagnosis and intervention in the RT planning.