Multi-view sparse attention network for glioma survival risk prediction.

BACKGROUND

Predicting the survival risk of gliomas is vital for personalized treatment plans. The latest survival risk prediction methods primarily rely on histopathology and genomics, which are invasive and costly. However, predicting survival risk based on non-invasive Magnetic Resonance Imaging (MRI) or handcrafted radiomics (HCRs) and clinical features has remained a challenge. The fusion of multi-view, non-invasive information holds the potential to improve risk prediction. Additionally, existing survival risk prediction methods typically depend on the Cox partial log-likelihood loss as their main optimization criterion, which may overlook the survival rankings among gliomas, leading to discrepancies between risk prediction and actual outcomes.

PURPOSE

This study aims to propose a non-invasive multi-view survival risk prediction network for gliomas to meet the clinical demand for efficient prognosis.

METHODS

This paper proposes a multi-view survival risk prediction network, which uses multi-view data as input, including 3D multi-modal MRIs, 2D images projected from MRIs, 1D HCRs features based on MRIs, and clinical information. In the feature encoder for each view, we design Pooling and Sparse Attention-based Transformer to extract risk-related features. We propose a Multi-View Complementary Attention Fusion module based on local and global attention to capture complementary features between different views and train a Cox model for survival risk prediction. We design a similarity loss based on cosine similarity to ensure the uniqueness of the extracted features between different views and design a pairwise ranking loss to address the Cox model's difficulty in discerning survival differences.

RESULTS

The experimental results demonstrate that our method performs well in glioma survival risk prediction. It achieves a C-index of 75.35% and 74.47% on the publicly available UCSF-PDGM and BraTS2020 datasets, surpassing other single-view and multi-view methods. Additionally, the proposed method has the lowest number of trainable parameters compared to other MRI-based methods, with only 29.07 million, achieving a trade-off between performance and parameter efficiency.

CONCLUSIONS

The proposed method highlights that we effectively fuse multi-view non-invasive information, offering advantages in survival risk prediction and advancing the research on glioma prognosis.