Glioma classification framework based on SE‐ResNeXt network and its optimization

Abstract

Aiming at the complexity of the glioma classification process, a classification framework based on SE-ResNeXt network is proposed to simplify the classification process of benign and malignant of gliomas. In addition, three optimization strategies are adopted to improve the accuracy. Firstly, the MultiStepLR strategy is used to adjust the learning rate dynamically in order to improve the learning ability of the network. Secondly, the one-hot label is optimized by the label smoothing strategy which can reduce the dependence of the network on the probability distribution of real labels and improve the prediction ability of the network. Finally, the transfer learning process is simplified by the transfer learning strategy on CE-MRI dataset, and the generalization ability of the network is improved. The experimental results show that the accuracy, sensitivity, specificity and AUC of the proposed method reach 97.45%, 98.35%, 94.93% and 0.9966 for the BraTS2017 dataset, 98.99%, 99.18%,98.33% and 0.9993 for the BraTS2019 dataset, respectively. Compared with the classical networks and other algorithms, the classification framework proposed in this paper has the best performance on glioma classification.