Multi-view sparse attention network for glioma survival risk prediction

Area of Science:

• Neuro-oncology
• Medical imaging analysis
• Machine learning in healthcare

Background:

• Glioma survival risk prediction is crucial for personalized treatment but current methods using histopathology and genomics are invasive and costly.
• Predicting glioma prognosis using non-invasive Magnetic Resonance Imaging (MRI) or handcrafted radiomics (HCRs) and clinical data remains challenging.
• Existing survival prediction models often use Cox partial log-likelihood loss, which may not effectively capture survival differences, impacting prediction accuracy.

Purpose of the Study:

• To develop a non-invasive, multi-view survival risk prediction network for gliomas to enhance prognostic efficiency.
• To address the limitations of invasive prediction methods by leveraging multi-modal, non-invasive data.

Main Methods:

• Proposed a multi-view survival risk prediction network integrating 3D/2D MRIs, handcrafted radiomics (HCRs), and clinical data.
• Employed a Pooling and Sparse Attention-based Transformer in feature encoders for risk-related feature extraction.
• Introduced a Multi-View Complementary Attention Fusion module for inter-view feature integration and utilized similarity and pairwise ranking losses for improved Cox model training.

Main Results:

• The proposed method achieved high performance in glioma survival risk prediction, with C-indices of 75.35% (UCSF-PDGM) and 74.47% (BraTS2020).
• Outperformed existing single-view and multi-view prediction methods.
• Achieved parameter efficiency with only 29.07 million trainable parameters, balancing performance and computational cost.

Conclusions:

• The study successfully demonstrated the effectiveness of fusing multi-view non-invasive information for glioma survival risk prediction.
• The proposed network offers significant advantages for clinical glioma prognosis, advancing non-invasive predictive modeling.