A comprehensive multi-domain dataset for mitotic figure detection

Area of Science:

• Digital pathology
• Computational biology
• Medical image analysis

Background:

• Mitotic figures in tumor tissue are key prognostic indicators.
• Automating mitotic figure detection is a significant research goal.
• Deep learning models struggle with domain shifts from varied data sources.

Purpose of the Study:

• Introduce the MIDOG++ dataset for mitotic figure detection.
• Address domain shift challenges in automated pathology.
• Enhance generalizability of deep learning models for tumor analysis.

Main Methods:

• Developed the MIDOG++ dataset with 11,937 mitotic figure labels across 503 histological specimens.
• Included seven diverse tumor types (e.g., breast carcinoma, lung carcinoma).
• Utilized specimens processed in multiple labs with various scanners to simulate domain shift.

Main Results:

• Evaluated state-of-the-art methods, confirming significant performance drops due to domain shifts in single-domain training.
• Demonstrated considerable improvement in model generalizability using a leave-one-domain-out approach.
• The MIDOG++ dataset is the first to encompass domain shifts from tumor types, labs, scanners, and species.

Conclusions:

• The MIDOG++ dataset provides a robust benchmark for evaluating domain generalization in mitotic figure detection.
• Addressing domain shift is critical for reliable deployment of AI in digital pathology.
• This dataset will facilitate the development of more robust and widely applicable automated pathology tools.