SMMILe enables accurate spatial quantification in digital pathology using multiple-instance learning.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in medicine

Background:

• Spatial quantification is crucial for computational pathology but often lost in multiple-instance learning (MIL) models.
• Existing MIL methods predict whole-slide image labels but lack spatial awareness, limiting clinical applications.
• Manual annotations are time-consuming and hinder scalability in pathology image analysis.

Purpose of the Study:

• To develop a computational pathology method that achieves accurate whole-slide image (WSI) prediction and superior spatial quantification.
• To mathematically prove that instance-level aggregation can improve spatial awareness in MIL models.
• To introduce and validate a superpatch-based measurable multiple-instance learning (SMMILe) method.

Main Methods:

• Developed a superpatch-based measurable multiple-instance learning (SMMILe) framework.
• Mathematically demonstrated the benefits of instance-level aggregation for spatial quantification in MIL.
• Evaluated SMMILe on 3,850 whole-slide images across 6 cancer types and 3 classification tasks.
• Benchmarked SMMILe against nine existing methods using two distinct encoders (ImageNet-pretrained and pathology-specific).

Main Results:

• SMMILe matched or exceeded state-of-the-art WSI classification performance across all evaluated tasks.
• The method demonstrated outstanding spatial quantification capabilities.
• Performance was consistent regardless of the encoder used (ImageNet or pathology-specific foundation model).

Conclusions:

• SMMILe effectively integrates spatial quantification with high-performance WSI classification.
• The proposed instance-level aggregation approach overcomes limitations of traditional MIL methods.
• SMMILe offers a powerful tool for advancing computational pathology and digital diagnostics.