부산대학교 도서관

High-Grade Glioma (HGG) is one of the major malignant tumors threatening human life. Accurate survival prediction of patients can help clinicians make better treatment decisions. However, the amount of HGG data is often small and insufficient, the related features extracted by a single feature extraction method are often incomplete, and the performance of a single classifier model is often poor, all of which will have a negative impact on the survival prediction of HGG. In this paper, deep features extracted from neural networks based on different tasks are combined with traditional radiomics features, and a multi-model integrated prediction method is proposed for survival prediction of HGG patients. Neural networks based on classification tasks and segmentation tasks are used to extract deep features, wherein neural networks based on classification tasks are used to extract information of category features and gray distribution and other related information in 2D image slices. The neural network based on 3D segmentation task is used to extract the spatial and shape information in the image. Its purpose is to explore deeper related information in limited HGG image data, so as to make full use of medical images. At the same time, we construct a multi-layer nested ensemble model composed of multiple weak classifiers by the concept of ensemble learning, which can accept the input of multi-source feature information to gain higher prediction accuracy in HGG dataset with small sample size and insufficient data. In the survival prediction of Br $a$ TS2020, the Accuracy score of our proposed method is 2.1% higher than that of other methods' with the best effect. The results show that this method has good prediction effect.