Imbalanced Multiclassification Challenges in Whole Slide Image: Cross-Patient Pseudo Bags Generation and Curriculum...

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in medicine

Background:

• Histopathological image multi-classification faces challenges with imbalanced datasets.
• Existing methods struggle with accurate classification under data imbalance.

Purpose of the Study:

• To propose a novel cross-patient pseudo-bag generation technique for imbalanced histopathological image classification.
• To address the challenge of distributional alignment in pseudo-bag generation.

Main Methods:

• Developed a cross-patient pseudo-bag generation framework extracting complementary pathological features.
• Implemented an affinity-driven curriculum contrastive learning strategy for distributional alignment.
• Pioneered multi-instance feature distribution mining, modeling inter-bag heterogeneity.

Main Results:

• Achieved significant performance improvements on three datasets with varying classification difficulties.
• Outperformed the second-best method by an average of 1.95% in F1 score and 2.07% in accuracy (ACC).

Conclusions:

• The proposed method effectively handles class imbalance in histopathological image classification.
• The framework offers a paradigm shift towards feature distribution mining for improved diagnostic accuracy.