Multi-instance curriculum learning for histopathology image classification with bias reduction.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology image analysis

Background:

• Multi-instance learning (MIL) excels at analyzing gigapixel histopathological images but faces challenges.
• Current MIL methods exhibit bias by focusing on easy instances and suffer from class imbalance, leading to false positives and skewed classification.

Purpose of the Study:

• To develop a multi-instance curriculum learning method to mitigate bias in histopathological image analysis.
• To improve classification performance and model interpretability in digital pathology.

Main Methods:

• Proposed a curriculum learning approach incorporating hard negative instance mining and positive instance augmentation using a diffusion model.
• Initialized MIL models with easy instances, then retrained with mined hard negatives and augmented positives via memory rehearsal.

Main Results:

• The proposed method effectively alleviates model bias inherent in traditional MIL approaches.
• Demonstrated improved classification performance and enhanced model interpretability in histopathological image analysis.

Conclusions:

• The novel curriculum learning strategy successfully addresses key limitations in MIL for histopathology.
• This approach offers a more robust and interpretable solution for analyzing complex digital pathology images.