Research On Prognosis Algorithm Of High Grade Glioma Based On Deep Learning

Glioma is one of the most common but hard to cure malignant neurological tumors.Accurate diagnosis for glioma patient is particularly important to improve treatment effect and survival rates.The World Health Organization grading system classifies glioma into grades Ⅰ-Ⅳ according to the degree of malignancy of tumor differentiation.Grade Ⅰ and Ⅱ belong to low grade glioma,and grade Ⅲ and Ⅳ belong to high grade glioma.In clinical practice,Magnetic Resonance Imaging(MRI)of patients is important basis for the glioma grading.Further studies showed that preoperative MRI images of patients with high grade glioma can be used for prognostic diagnosis and guide patients to follow-up in time.During these years,deep learning has been extensively used in the field of medical imaging.However,studies on the prognosis of high grade glioma based on preoperative MRI images have found that the accuracy is still far from clinical requirement.First,the training of deep learning models relies on the support of multicenter data,and the privacy protection requirement of medical data leads to the serious data prison.The traditional data centralized model training method is limited by the amount of data,which limit the performance improvement of glioma grading algorithm.Second,existing high grade glioma prognostic algorithms directly analyze MRI images that have only undergone conventional preprocessing,which may make feature overlap and make it difficult for subsequent models to extract effective features.In view of the above problems,this paper conducts algorithm research on the two stages of glioma grading and high grade glioma prognosis.First,an improved glioma grading algorithm based on deep learning using federated learning is proposed.The network with attention mechanism is designed to extracting features in the region of interest(ROI)through spatial and channel angle weighting,which provides a model basis for research.In order to improve the imbalanced data distribution of each client,the federation algorithm based on dynamic weighted aggregation of loss values is proposed.The designed algorithm can complete the model training only by sharing the weight of model from each client without revealing the original data such as user information.Simulation experiments in different data prison scenarios show that the improved Fed WL algorithm has a 0.65%-6.06% improvement in accuracy compared with the traditional federated learning Fed AVG algorithm,and at least 0.73%-2.74%improvement compared with other distributed training algorithms.Subsequently,dedicated to realizing the prognosis prediction of high grade glioma patients through preoperative MRI images,this paper proposes a high grade glioma prognosis algorithm based on deep convolutional neural network.In this algorithm,the original image is fitted with a double Gaussian mixture model to obtain the Habitat Map.The Habitat Map,the corresponding segmented image and the patient information are used as input.The data classification model HabitatXGBoost composed of convolutional neural network Habitat Net and feature classification network XGBoost is designed to effectively alleviate the problem of overfitting.The results show that the proposed HabitatXGBoost algorithm has a 7.08%-21.93% improvement in accuracy compared with other deep learning and traditional machine learning methods.The work of this paper is expected to provide a clinical reference for the prognosis diagnosis of high grade glioma based on medical imaging.