Label Cleaning Multiple Instance Learning: Refining Coarse Annotations on Single Whole-Slide Images.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Accurate annotation of whole-slide images (WSIs) is crucial for cancer diagnosis and research.
• Manual annotation is time-consuming, costly, and requires expertise.
• Coarse annotations are easier but lack precision for detailed analysis.

Purpose of the Study:

• To develop a method for refining coarse annotations in digital pathology.
• To address the limitations of existing methods that require large datasets for training.
• To improve the accuracy of segmentation from imprecise labels.

Main Methods:

• Introduced Label Cleaning Multiple Instance Learning (LC-MIL).
• Processed patches from a single WSI with inaccurate labels jointly.
• Utilized a multiple instance learning framework to mitigate label impact and refine segmentation.

Main Results:

• LC-MIL significantly refines coarse annotations across diverse cancer types (breast, liver, colorectal).
• Outperformed state-of-the-art methods, even when learning from a single slide.
• Demonstrated effective refinement of real pathologist-drawn annotations.

Conclusions:

• LC-MIL is a promising, lightweight tool for improving annotation granularity.
• Enables efficient generation of fine-grained annotations from coarse pathology datasets.
• Reduces the burden on pathologists while enhancing data quality for research and AI.