Rethinking mitosis detection: Towards diverse data and feature representation for better domain generalization

Area of Science:

• Computational pathology
• Digital pathology
• Medical image analysis

Background:

• Mitosis detection is crucial in computational pathology but challenging due to cell heterogeneity.
• Current methods often increase model complexity, risking overfitting and limiting generalizability.
• Biological knowledge and feature diversity are often overlooked in existing approaches.

Purpose of the Study:

• To develop a generalizable mitosis detection framework by addressing data and feature diversity.
• To improve the robustness and applicability of mitosis detection models in computational pathology.
• To systematically investigate the impact of morphological appearances and data balancing on detection performance.

Main Methods:

• Proposed a novel generalizable framework (MitDet) for mitosis detection.
• Implemented diversity-guided sample balancing (DGSB) for data diversity.
• Utilized inter- and intra- class feature diversity-preserved module (InCDP) for feature diversity.
• Introduced a Stain Enhancement (SE) module to boost domain-relevant diversity.

Main Results:

• MitDet significantly outperformed state-of-the-art (SOTA) methods on multiple mitosis detection datasets.
• The model demonstrated superior generalizability on both internal and unseen test sets.
• Ablation studies confirmed the effectiveness of balancing data and feature diversity.

Conclusions:

• The proposed MitDet framework achieves SOTA performance in mitosis detection.
• Balancing data and feature diversity is key to improving generalizability in computational pathology.
• This work offers a new perspective for future research in mitosis detection and computational pathology.