Automated mitosis detection in stained histopathological images using Faster R-CNN and stain techniques

Area of Science:

• Digital pathology
• Computational oncology
• Artificial intelligence in medicine

Background:

• Accurate mitosis detection is crucial for cancer diagnosis and treatment planning.
• Manual counting of mitotic figures by pathologists is labor-intensive and prone to errors.
• Deep learning (DL) offers a promising avenue for automating mitosis detection in histopathological images.

Purpose of the Study:

• To investigate the efficacy of DL object detection models for automated mitosis detection in stained histopathological images.
• To address the challenge of domain shift in histopathological image analysis using stain augmentation and normalization.
• To compare the performance of a two-stage Faster R-CNN model against one-stage RetinaNet models.

Main Methods:

• Development and implementation of a two-stage Faster R-CNN object detection model.
• Application of stain augmentation and normalization techniques to mitigate domain shift.
• Experimental validation using the MIDOG++ dataset, a recent benchmark for mitosis detection.
• Comparison with previously developed one-stage RetinaNet frameworks.

Main Results:

• The Faster R-CNN model, incorporating stain techniques, achieved highly accurate and reliable mitosis detection.
• Favorable F1-scores were observed across diverse scenarios and tumor types, validating the DL models' effectiveness.
• RetinaNet models demonstrated faster processing speeds while maintaining competitive performance.
• The study underscored the critical role of addressing domain shifts and mitotic figure counts for robust diagnostic tools.

Conclusions:

• Deep learning object detection models, particularly Faster R-CNN with stain normalization, significantly enhance automated mitosis detection accuracy in histopathology.
• The findings support the development of more reliable and efficient computational tools for cancer diagnosis.
• Effective handling of domain shift is essential for the generalizability and robustness of automated pathology systems.