| Diffuse large B lymphoma is a long-standing disease that has a high incidence,especially in China,and poses a significant health risk to our population.Early identification of high-risk groups will play a crucial role in the selection of future treatment options and prognosis of patients with this disease.Currently,the clinical treatment of this condition cannot be achieved without the determination of the lesion,but currently the PET(Positron Emission Tomography)images are generally sketched manually in order to target the area of the lesion,which is time-consuming and unpredictable,and has a significant impact on the diagnosis of the condition.This approach is inevitably time-consuming and unpredictable,and can have a significant impact on the diagnosis of lesions.After the lesion is outlined,feature extraction is needed,and the most used feature extraction technology is imaging histology technology,which can be used to extract specific features of tumors,and how to accurately and quickly understand the heterogeneity of tumors becomes another important direction for effective treatment,which also indicates a direction for the prognosis of tumor patients.In order to investigate how to achieve rapid localization and size determination of tumors has become a difficult problem,and how to achieve accurate judgment of patient prognosis is also a problem,which is studied in thesis.In thesis,we propose a 2D PET automatic segmentation algorithm based on deep learning to address the problems of inefficiency and inconsistent judgment criteria caused by the traditional software manual outlining of lesion areas.First,we use the slice data from medical institutions that have already outlined the lesion areas to divide the dataset;for the characteristics of uneven distribution of diffuse large B lymphoma lesion areas,the characteristics of different lesion sizes in the slice,and the lack of boundary conditions in the PET image itself,we build a deep learning-based 2D PET image automatic segmentation model,and use the model to train the dataset;After that,the parameters obtained from the training are used to predict the validation set and calculate the corresponding evaluation indexes.The experiments show that the 2D PET image segmentation model has different degrees of improvement in each index: the Dice value reaches0.8189 and the specificity value PPV reaches 0.8633,which indicates that the model can play a better role in the future diagnosis of the disease.Based on the automatic segmentation technique to obtain pathological features,thesis proposes a variety of machine learning automatic feature selectors based on ensemble learning methods.First,using the data obtained by automatic segmentation of 2D PET images,feature information of 371 groups of patients was extracted using a batch processing method,and an automatic feature selector based on multiple machine learning methods with integrated learning approach was used to select key features based on cardinality distribution and mutual information methods,so as to construct a feature signature with stable and better predictive ability in risk stratification of patients.The factors at the imaging histology level and clinical level are fully considered to improve the identification of high-risk patients with poor prognosis.The experiments showed that the optimized combined model outperformed the other two types of models in terms of calibration,net gain and consistency indexes(0.733 in the PFS training set and 0.728 in the validation set;0.743 in the OS training set and 0.731 in the validation set)for the prediction of PFS and OS,further validating the prognostic value of imaging histology features for cancer patients,and proving that the The proposed combined model can be of greater benefit in the clinical setting. |
PDF Full Text Request