AdvMIL: Adversarial multiple instance learning for the survival analysis on whole-slide images.

Area of Science:

• Computational pathology
• Medical image analysis
• Deep learning for survival analysis

Background:

• Survival analysis of whole-slide images (WSIs) is crucial for patient prognosis.
• Existing weakly-supervised deep learning models for WSIs are limited by classical survival analysis rules and require fully-supervised learning on small labeled datasets.
• Current models provide point estimations of time-to-event and struggle with utilizing unlabeled data.

Purpose of the Study:

• To propose a novel adversarial multiple instance learning (AdvMIL) framework for WSI survival analysis.
• To enhance survival distribution estimation and enable semi-supervised learning capabilities.
• To improve the robustness of survival analysis models against image noise and data corruption.

Main Methods:

• Developed an adversarial multiple instance learning (AdvMIL) framework integrating adversarial time-to-event modeling with multiple instance learning (MIL).
• Designed AdvMIL as a plug-and-play module to upgrade existing MIL-based end-to-end methods.
• Utilized extensive experiments to evaluate the framework's performance and capabilities.

Main Results:

• AdvMIL significantly improved the performance of mainstream WSI survival analysis methods at a low computational cost.
• The framework effectively enabled the utilization of unlabeled data through semi-supervised learning.
• AdvMIL demonstrated enhanced model robustness against patch occlusion and common image noises.

Conclusions:

• The proposed AdvMIL framework offers a novel adversarial MIL paradigm for survival analysis in computational pathology.
• AdvMIL promotes the development of more accurate, robust, and data-efficient survival analysis models for WSIs.
• This framework has the potential to advance prognostic estimation and clinical decision-making in digital pathology.