In International journal of computer assisted radiology and surgery
PURPOSE : As the most common primary intracranial tumor, glioblastoma (GBM) is a malignant tumor that originated from neuroepithelial tissue, accounting for 40-50% of brain tumors. Precise survival prediction for patients suffering from GBM can not only help patients and doctors formulate treatment plans, but also help researchers understand the development of the disease and stimulate medical development.
METHODS : In view of the tedious process of manual feature extraction and selection in traditional radiomics, we propose an end-to-end survival prediction model based on DenseNet to extract the features of magnetic resonance images including T1-weighted post-contrast images and T2-weighted images through two-branch networks. After segmenting the region of interest, the original image, the image of tumor region and the image without tumor are combined as input sample sets with three channels. Additionally, for some patients having only one of T1- or T2-weighted images, One2One CycleGAN is used to generate the T1 image from the T2 image, or vice versa. Flipping and rotating are also used for sample augmentation.
RESULT : By using the augmented training sample set to train the model, the classification and prediction accuracy of the two-branch DenseNet survival prediction model can reach up to 94%, and the Kaplan-Meier survival curve indicates that the model can classify patients into high-risk group and low-risk group based on whether they could survive for more than three years.
CONCLUSION : The classification and prediction results of the model and the survival analysis demonstrate that our model can get superior classification results which can be referenced by doctors and patients' families for developing medical plans. However, improving the loss function and expanding the sample size can further improve the prediction results, which are the target of our subsequent research.
Fu Xue, Chen Chunxiao, Li Dongsheng
Deep learning, Generative adversarial networks, Glioblastoma, Radiomics, Survival prediction