A universal multiple instance learning framework for whole slide image analysis.

Area of Science:

• Computational pathology
• Digital pathology
• Medical image analysis

Background:

• Digital whole slide images (WSI) have advanced computational pathology.
• Patch-level annotation of WSIs is difficult and time-consuming due to high resolution.
• Fully supervised methods are limited by the challenges of WSI annotation.

Purpose of the Study:

• To develop a universal framework for weakly supervised WSI analysis.
• To overcome the limitations of patch-level annotation requirements.
• To improve the accuracy of WSI classification and lesion detection.

Main Methods:

• A Multiple Instance Learning (MIL) framework is proposed for weakly supervised WSI analysis.
• A multi-dimensional feature aggregation module considers feature distribution, instance correlation, and instance-level evaluation.
• Key components include instance-level standardization, deep projection, self-attention, instance-level pseudo-labeling, and a key instance selection module.

Main Results:

• The proposed method achieved competitive performance on multiple benchmark datasets (Camelyon16, TCGA-NSCLC, SICAPv2, PANDA).
• A maximum improvement of 14.6% in classification accuracy was observed compared to recent methods.
• The framework effectively enhances WSI prediction accuracy through improved feature aggregation and instance selection.

Conclusions:

• The developed method improves whole slide image classification accuracy using weak supervision.
• The framework enables more accurate detection of lesion areas within WSIs.
• This approach addresses the annotation bottleneck in computational pathology.