MILA-MIL:Mamba-Inspired Linear Attention Multiple Instance Learning for Whole-Slide Image Survival Prediction

Area of Science:

• Computational pathology
• Precision oncology
• Artificial intelligence in medicine

Background:

• Survival prediction from Whole-Slide Images (WSIs) is crucial for precision oncology but challenged by gigapixel image scale and tumor heterogeneity.
• Existing Multiple Instance Learning (MIL) methods struggle to balance local morphological details with long-range prognostic dependencies.

Purpose of the Study:

• To develop a novel dual-branch framework, MILA-MIL, to integrate local micro-anatomical features and global survival information for improved prognostic prediction.
• To address the limitations of current MIL approaches in capturing both fine-grained pathological gradients and scalable long-range dependencies.

Main Methods:

• Proposed a novel dual-branch framework, MILA-MIL, incorporating a Pinwheel Convolution (P-Conv) module for directional morphological gradients and a Mamba-inspired Linear Attention (MILA) branch for efficient global context modeling.
• Utilized a gated fusion mechanism to dynamically integrate local and global representations.
• Evaluated the framework on six diverse cancer cohorts.

Main Results:

• MILA-MIL achieved state-of-the-art performance across six cancer cohorts, demonstrating superior stability and predictive power compared to existing MIL aggregators.
• The framework effectively synergizes directional morphology with scalable global modeling for enhanced prognostic accuracy.

Conclusions:

• MILA-MIL offers a robust and interpretable solution for computational pathology, bridging local and global information for survival prediction.
• This approach has significant potential to enhance clinical decision-making in personalized cancer management.