Prediction Of Radiotherapy Prognosis For Head And Neck Cancer By Three-Dimensional Dose Distribution

Head and neck cancer(HNC)is the sixth most common cancer in the world.Most patients are locally advanced when diagnosed,and the 5-year survival rate is less than 50%.Distant metastasis and normal tissue complications due to radiation-induced toxicities are the main reasons for the poor prognosis of HNC.Therefore,accurately predicting the prognosis of HNC patients and assessing the risk of tumors before treatment will help doctors better develop personalized treatment plans for patients.CT imaging has been widely used in the diagnosis,staging,and prognosis evaluation of head and neck cancer because of its ability to reflect information about the internal tissue structure,lesion shape,size,location and density of patients.The model based on medical images only includes the patient’s intrinsic anatomical structure or metabolic information in the pre-treatment stage,ignoring the heterogeneity of the patient’s treatment plan.Three-dimensional(3D)dose distribution in radiotherapy planning contains important treatment information and is easy to obtain.However,there are few studies on the prognostic value of 3D dose distributions in HNC.Therefore,aiming at the distant metastasis of head and neck cancer and normal tissue complications caused by radiation-induced toxicities,this study applied radiomics and deep learning strategies to 3D dose distribution and CT images,and constructed a multimodal fusion model,trying to improve the accuracy of the prediction model.In the prediction of radiation-induced temporal lobe injury(TLI)in head and neck cancer,CT images and 3D dose distribution were incorporated into the prediction of TLI with the help of radiomics method,the prognostic factors of TLI were explored,and a TLI prediction model was finally established.The results showed that the prediction ability of CT radiomics features was weak,while the prediction ability of dosiomics features was higher than that of CT radiomics features,D1cc(C-index in the prospective test and external test cohort respectively:0.79 vs.0.55 vs.0.70,0.77 vs.0.52 vs.0.72).The dosiomics risk model combining dosiomics signature,D1cc,and age showed higher accuracy,significantly outperforming the QUANTEC model of Dmax and the Wen’s model of D0.5cc combined with age and T stage(C-index in the prospective test and external test cohort respectively:0.81 vs.0.54 vs.0.76,0.79 vs.0.53 vs.0.75).In the prediction of distant metastasis(DM)in head and neck cancer,the 3D network 3D-SEResNet10 was used to construct an end-to-end DM prediction model,and the prognostic value of CT images and 3D dose distribution in DM of head and neck cancer was explored.In this study,CT images or 3D dose distribution were combined with the corresponding GTV contour to construct CT+GTV and dose+GTV models(named CG and DG model),respectively.In addition,we further combined CT images,3D dose distribution and GTV contour to construct a CT+ dose+GTV model(named CDG model),expecting to establish a prediction model with higher performance,and compared it with the constructed CG model and DG models.The results showed that the prediction ability of 3D dose distribution was similar to that of CT image in the prediction of DM(AUC:0.93 vs.0.93,C-index:0.88 vs.0.89).CT images and 3D dose distribution provide complementary information at different levels.By combining these information,the model performance is further improved(AUC:0.96,C-index:0.91).In summary,the 3D dose distribution has good prognostic value and robustness in HNC prognosis,showing good predictive ability under different prognosis conditions of HNC and different data mining methods.