Hard Negative Sample Mining for Whole Slide Image Classification.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in medicine

Background:

• Weakly supervised whole slide image (WSI) classification presents challenges due to absent patch-level labels and significant computational demands.
• Current methods often rely on self-supervised patch-wise feature representations within multiple instance learning (MIL) frameworks.
• Existing fine-tuning approaches using pseudo-labeling primarily focus on high-quality positive patch selection.

Purpose of the Study:

• To enhance feature representations and decrease computational costs in weakly supervised WSI classification.
• To introduce a strategy for mining hard negative samples during the fine-tuning process.
• To develop a novel patch-wise ranking loss function for improved MIL performance.

Main Methods:

• Implementation of hard negative sample mining during the fine-tuning stage of feature representation.
• Development and application of a novel patch-wise ranking loss function tailored for MIL.
• Validation of the proposed methods on two publicly available whole slide image datasets.

Main Results:

• Demonstrated improvement in feature representations through hard negative sample mining.
• Significant reduction in the overall training cost for WSI classification tasks.
• Validation of the proposed patch-wise ranking loss in enhancing MIL performance, as shown by experimental results.

Conclusions:

• The proposed approach of mining hard negative samples effectively improves feature representations and reduces computational costs in weakly supervised WSI classification.
• The novel patch-wise ranking loss function offers a superior method for exploiting hard negative samples within MIL frameworks.
• The findings suggest a promising direction for more efficient and effective computational pathology analysis.