Intervention-Based Mixup Augmentation for Multiple Instance Learning in Whole-Slide Image Survival Analysis.

Area of Science:

• Computational pathology
• Medical image analysis
• Machine learning

Background:

• Multiple instance learning (MIL) is crucial for analyzing gigapixel whole-slide images (WSIs).
• Data augmentation is vital for addressing data scarcity in survival analyses.
• Existing augmentation methods are insufficient for multi-instance regression in survival analysis due to label quality and data imbalance challenges.

Purpose of the Study:

• To introduce InterMix, a novel plug-and-play multi-instance data augmentation scheme.
• To address the challenges of high-quality pseudo-labeling and long-tail data distribution in survival analysis.
• To improve the performance of computational pathology models in survival prediction.

Main Methods:

• InterMix utilizes an intervention-inspired approach for data augmentation.
• Sub-bag assessment rules are employed to ensure high-quality pseudo-labels for augmented samples.
• A complementary mix strategy is proposed to generate diverse minority-augmented instances, addressing the long-tail distribution.

Main Results:

• InterMix demonstrated superior performance across three independent datasets.
• The method effectively handles data scarcity and imbalanced distributions in survival analysis.
• InterMix improves model focus on critical regions relevant to pathologists.

Conclusions:

• InterMix offers a significant advancement in data augmentation for multi-instance survival analysis.
• The proposed techniques effectively overcome key limitations of existing methods.
• InterMix has the potential to enhance diagnostic accuracy and efficiency in computational pathology.