| Nasopharyngeal carcinoma is one of the most common malignant tumors in our country,and it ranks first in the incidence of head and neck malignant tumors in our country.Early treatment is one of the most effective ways to improve the efficacy of cancer treatment.Both the early diagnosis of primary nasopharyngeal carcinoma and the early diagnosis of recurrent nasopharyngeal carcinoma can win an earlier treatment opportunity for patients and improve the cure rate and survival rate of patients to a certain extent.Because the site of nasopharyngeal cancer is relatively hidden and the clinical symptoms in the early stage are not typical and lack specificity,it is difficult to diagnose early stage nasopharyngeal carcinoma,leading to missed diagnosis and misdiagnosis.Benign hyperplasia of the nasopharyngeal wall and adenoids is very common in adults,and its imaging features are similar to those of early nasopharyngeal carcinoma,which increases the difficulty of early nasopharyngeal carcinoma diagnosis.Diagnosis of local tumor recurrence is often influenced by changes in nasopharyngeal tissue after treatment,such as necrosis,edema,inflammation,fibrosis,and scarring.Due to the high spatial resolution of soft tissue examination and the sensitivity of bone marrow infiltration,MRI is the preferred imaging modality for nasopharyngeal carcinoma identification,staging,efficacy evaluation,and post-treatment follow-up in clinical practice.Computer-aided diagnosis technology based on deep learning provides more scientific and objective results for cancer diagnosis,which can shorten the time of diagnosis,improve the accuracy of cancer diagnosis,and reduce the rate of missed diagnosis.Therefore,this study attempted to combine the magnetic resonance image of nasopharyngeal carcinoma and deep learning methods to efficiently and accurately diagnose patients with early nasopharyngeal carcinoma and patients with recurrent nasopharyngeal carcinoma,so as to strive for earlier treatment opportunities for patients and prolong the length of patient’s life and improve patients’ quality of life.Based on the deep learning method,this thesis has established corresponding diagnostic models for early nasopharyngeal carcinoma image data and post-treatment nasopharyngeal carcinoma image data.The main research contents are as follows:(1)For early nasopharyngeal carcinoma image data,this thesis designed a convolutional neural network model based on transfer learning to assist in the diagnosis of early nasopharyngeal carcinoma and benign hyperplasia patients.To extract features that are more useful for model classification,we delineated regions of interest on the image data.By adopting the five-fold cross-validation method to divide the data set and using data augmentation technology to expand the training data set,the generalization ability and robustness of the model were improved.The pre-trained DenseNet169 model achieved the best results,with an accuracy of 0.9296,a sensitivity of 0.9452,a specificity of 0.9066,an F1 score of 0.9379,and an AUC value of 0.9712.(2)For the post-treatment nasopharyngeal carcinoma image data,this thesis proposed a new convolutional neural network model to diagnose patients with recurrent nasopharyngeal carcinoma and post-radiotherapy changes.We added a coordinate attention module to the improved Inception V2 module,and then built a new model.The model can extract multi-scale features and channel attention features with direction-aware and position-aware information,which enhanced the feature expression ability of the model.Furthermore,to improve the robustness of the results,the average results of the fifty hold-out method were applied to evaluate the performance of the model.The results showed that our constructed CNN model achieved an accuracy of 0.8608,a sensitivity of 0.8,a specificity of 0.8979,an F1 score of 0.8136,and an AUC value of 0.8949.Compared with other models,the new CNN model was more accurate and takes less time. |
PDF Full Text Request