MPSurv: End-to-End Multi-model Pseudo-Label Model for Brain Tumor Survival Prediction with Population Information Integration

Predicting brain tumor survival can aid physicians in better assessing the efficacy of treatments and adjusting treatment plans in clinical practices to enhance patient survival. Recently, deep learning techniques have attracted massive attention in predicting brain tumor survival. However, the majority of existing methods necessitate at least two or more independent networks for knowledge sharing later in the model and overlook the significance of population information. In this paper, we propose an end-to-end multi-model brain tumor survival prediction (MPSurv) model that incorporates patient population information. Moreover, given the presence of censored data, we propose to address this issue by generating pseudo-labels, which in turn augments the original data and improves the utilization of the dataset. We have collected and supplemented survival labels based on the BraTS 2021 dataset for the training and validation of segmentation and prediction tasks. Experimental results demonstrate that our model enhances the accuracy of brain tumor survival prediction and exhibits superior generalizability. The source code is available at: https://github.com/APTX574/MPSurv.