Mitosis Counting in Breast Cancer: Object-Level Interobserver Agreement and Comparison to an Automatic Method

Area of Science:

• Histopathology
• Computational Pathology
• Breast Cancer Research

Background:

• Mitosis counting in breast cancer histology is crucial but subjective and lacks reproducibility.
• Current methods focus on overall counts, not individual object agreement, limiting insight.

Purpose of the Study:

• To analyze object-level interobserver agreement in mitosis counting.
• To develop and evaluate an automatic mitosis detection method robust to staining variability.
• To compare automatic method performance against expert observers on external datasets.

Main Methods:

• Conducted an object-level interobserver agreement study with three pathologists.
• Developed an automatic mitosis detection algorithm trained on diverse histopathology images.
• Validated the automatic method on an external dataset from different pathology labs.

Main Results:

• Pathologists frequently disagreed on individual mitosis identifications, even when overall counts agreed.
• Disagreement increased for smaller objects, suggesting size constraints could improve protocols.
• The automatic method demonstrated unbiased mitosis counting with substantial agreement with human experts.

Conclusions:

• Object-level analysis reveals significant interobserver variability in mitosis counting.
• Automatic mitosis detection offers a reproducible and unbiased approach to this critical biomarker assessment.
• Further refinement of mitosis counting protocols, potentially including size constraints, is warranted.