Gliomas are common tumors with extremely high malignancy in the central nervous system. Due to their strong heterogeneity, traditional survival prediction methods often overlook the correlations among the “tumor-edema-necrosis” multi-lesions and the deep fusion of multimodal data, making it difficult to meet the clinical demand for accurate prognostic assessment. This study proposes an integrated framework combining the Efficient Attention UNet (EA-UNet) segmentation network, as well as the supervised and unsupervised learning modules that include Factorization Machine (FM) and Cross-Attention (Cross-A), for glioma segmentation and survival prediction. EA-UNet shows excellent segmentation performance on the BraTS2021 dataset, with an average Dice coefficient of 0.8546 and only 0.4840M parameters. After fusing MRI images, clinical and genetic multimodal data via FM and Cross-A, the full-modal model achieves a maximum C-index of 0.736 (on the UCSF PDGM dataset). Performance decreases significantly when key modules are removed, and the model outperforms traditional methods and other deep learning models overall. Hazard score classification also confirms its effectiveness in aiding clinical decision-making. In conclusion, this multimodal integration framework performs prominently in glioma segmentation and survival prediction, providing reliable technical support for clinical accurate prognosis assessment and personalized treatment planning, with strong clinical application potential. • Proposes lightweight efficient Attention UNet for glioma segmentation and survival prediction • Integrates Factorization Machine and Cross-Attention modules to fuse multimodal data • Outperforms traditional methods and other deep learning models in glioma-related tasks.
Guo et al. (Sun,) studied this question.