Iterative multiple instance learning for weakly annotated whole slide image classification.

Area of Science:

• Computational pathology
• Digital histopathology
• Machine learning in medicine

Background:

• Whole slide images (WSIs) are vital for histopathology but their high resolution complicates detailed annotation.
• Classifying WSIs often uses multiple instance learning (MIL), treating the WSI as a bag of instances (patches).

Purpose of the Study:

• To develop a novel iterative multiple instance learning (IMIL) method for classifying WSIs using only slide-level labels.
• To enhance the accuracy of histopathological analysis through improved WSI classification.

Main Methods:

• Proposed an iterative MIL (IMIL) method that collaboratively learns instance and bag representations.
• Implemented iterative fine-tuning of a feature extractor using selected instances and pseudo-labels from attention-based MIL pooling.
• Employed self-supervised learning for feature extractor initialization, attention score-based sample selection, and confidence-aware loss for robust training.

Main Results:

• IMIL-SimCLR achieved optimal classification performance on Camelyon16 and KingMed-Lung datasets, outperforming the CLAM baseline by up to 4.25% in average AUC.
• IMIL-ImageNet demonstrated superior performance on the TCGA-Lung dataset with 96.55% AUC and 96.76% accuracy, surpassing CLAM by 1.65% AUC and 2.09% accuracy.
• The IMIL method proved effective across public and in-house datasets for various WSI classification tasks.

Conclusions:

• The proposed iterative MIL (IMIL) method offers a significant advancement in WSI classification for histopathology.
• IMIL demonstrates superior performance compared to state-of-the-art MIL methods across multiple datasets and classification tasks.
• This method holds promise for improving diagnostic accuracy and efficiency in digital pathology.