| Head and neck cancer(H&N)is a common malignancy in the world,and locoregional recurrences(LR)is one of the main causes of poor prognosis for H&N Assessing tumor risk and selecting an individualized treatment schedule before treatment is critical to improving patient prognosis.In clinical practice,the indicators including patient's age,tumor stage and so on are commonly used to assess tumor risk Furthermore,in recent years with the development of medical imaging technology,some studies revealed that the digital information in medical imaging can decode the characteristics of the tumor,which was a potential source of information for tumor risk assessment.Other studies showed that the dose-volume histogram produced during the stage of radiotherapy planning also had certain prognostic ability,so the effect of dose distribution quality on the prognosis of patients receiving radiotherapy should be fully considered in prognosis study.But the existing prognostic studies only involved two-dimensional dose information.By contrast,the three-dimensional(3D)dose distribution contains more dose information,however,few studies explored the prognostic value of the 3D dose distribution.Therefore,this paper attempted to extract quantified dose features from the 3D dose distribution of H&N cancer,then found out the dose factors affecting the LR of H&N cancer and established the LR prognostic model.Feature mining adopted two methods,including the traditional radiomics method and the deep learning methodThe radiomics strategy was applied in the 3D dose distribution of tumor area to extract a number of the first order and texture features,then the model describing the correlation of dose features and LR was established using Cox proportional hazard regression.This model was further compared with the model based on clinical parameters such as age and tumor stage,and the models based on CT and PET images.Meanwhile,we also combined dose+ images and dose+images+clinical parameters,attempted to improve the performance of the LR prediction models.237 H&N cancer cases from four treatment institutions were collected from an open database,wherein,141 cases from two institutions were selected for model training,and 96 cases from the other two institutions were selected for model evaluation.The results showed that the C-index of the model established by dose distribution was 0.60,which was higher than CT model(0.54),PET model(0.59)and CT+PET model(0.58),and was equivalent with the clinical model(0.60);the combination of dose and medical images could improve the performance of models(0.66),that is 3D dose distribution and medical image are complementary in the prediction of LR;but the further combination of clinical indicators was not conducive to improving model performance(0.56).Besides,a statistical analysis showed that patients with low risk of LR of H&N cancer had a smaller dose fluctuation range and a higher relative lowest doseThe features extracted by the traditional radiomics method are artificial,which may exclude some important image information.Therefore,we utilized the survival convolutional neural network(SCNN)of deep learning to automatically extract the features with high LR predictive power from dose distribution and images for establishing the LR prediction model.The results indicated again the complementarity of dose distribution and medical images in the prediction of recurrence,that is the integration of dose and medical images could further improve the performance of models.This finding illustrated that the prognostic value of 3D dose distribution for H&N LR was robust when using different modeling methods.Furthermore,compared with the traditional radiomics-based model,the SCNN-based model had higher performance.In summary,two image analysis methods were used in this paper to demonstrated the LR predictive power of 3D dose distribution in head and neck cancer. |
PDF Full Text Request