Artificial Intelligence-Based Automatic Mitosis Scoring in Breast Cancer Improves Inter-Observer Concordance and Efficiency.

Abstract

Background: Evaluation of mitotic activity is crucial for breast cancer treatment; however, manual scoring is imprecise, with considerable inter-observer variation, but comprehensive image analysis may address this problem. We designed an artificial intelligence-based workflow for automatic mitosis scoring of breast cancer images.

Materials and methods: Ninety-three cases, including 117 whole-slide images, were enrolled and 97 images were used to train the mitosis detection model. The annotation of the first dataset was guided by phosphohistone-H3 immunohistochemical restraining to build an auxiliary label model. The second dataset was expanded to include interactive learning. The automatic scoring framework included image partitioning, epithelial area segmentation, mitosis detection, hotspot analysis, and score classification. Clinical utility was evaluated by 3 pathologists using 20 slide images.

Results: The mitosis model achieved an F1-score of 0.75, which is comparable to that of other state-of-the-art algorithms. The automatic scoring workflow located hotspots by thoroughly analyzing the entire slide, which improved inter-pathologist concordance, but slightly underestimated the mitosis score. In addition, the reading time decreased significantly from an average of 452 s to 52 s for each case (P < .01).

Conclusion: This study demonstrated that an integrated automatic mitosis scoring system could boost pathology end users to efficiently achieve higher concordance, providing a solid foundation for precision medicine. Novel methodologies are essential in assessing mitotic activity in the era of digital pathology.