At present, mitosis detection in breast histopathology images is a critical issue for breast cancer grading. Due to the breast tissue having a complex structure, and mitosis and non-mitosis cells being similar to each other, traditional methods for detecting mitosis cells are not responsive. This study proposes a high F1-score mitosis detection approach using an improved Faster R-CNN with a Self-Attention mechanism. This mechanism allows the model to focus on the key feature that distinguishes between mitosis and non-mitosis cells and reduces irrelevant background elements. The proposed approach improves mitosis detection in sensitive and challenging conditions. This method performed well in detecting diverse mitosis shapes that are unique to each tissue, unlike conventional methods, and provides a reliable tool for pathologists in mitosis diagnosis and grading of breast tumors. The proposed method not only improves the F1-score of automated analysis of breast histopathology images, but also can help in making correct clinical decisions in breast cancer management. We evaluated our method using two public datasets: ICPR 2012 and ICPR MITOS-ATYPIA Challenge 2014. The results show that the F1-score of the proposed approach outperforms other algorithms in this field. On the combined dataset, we achieved an F1-score of 95.31%, a Sensitivity of 94.43%, and a precision of 96.22%.
Ayashm et al. (Thu,) studied this question.