by Christos Evangelou, MSc, PhD – Medical Writer and Editor

The identification and classification of mitotic figures are critical for the diagnosis and grading of tumors. However, pathologists often face challenges in obtaining consistent results. The 1000 Mitoses Project, an international collaborative study aimed at setting new standards for mitotic figure classification, brought together pathologists from around the globe to create a consensus-based approach.

Published in the International Journal of Surgical Pathology, the results of this study paved the way for more accurate and reproducible cancer diagnostics.

Addressing a Critical Need

Mitotic figures — cells in the process of division — are commonly used by pathologists to determine the aggressiveness of tumors. Accurate identification and counting of these figures can directly influence the diagnosis, treatment decisions, and prognosis of patients with cancer. Nevertheless, identifying and counting mitotic figures is subjective and prone to variability, leading to inconsistencies in diagnosis and treatment decision making. The 1000 Mitoses Project was conducted to address this issue, aiming to harmonize and improve the accuracy of mitotic figure identification across the pathology community.

“The motivation for conducting the study stemmed from the realization of inconsistencies in the interpretation of mitotic figures among pathologists, despite their critical role in tumor diagnosis, grading, and classification. Given the lack of comprehensive data on this subject and the potential to foster international collaboration through publicly accessible datasets and social media, the study aimed to create a robust dataset of mitotic figures to aid in future training of AI tools in pathology,”

said the corresponding author Matthew Cecchini, MD, PhD. Dr. Cecchini is a pathologist at the London Health Sciences Centre and an assistant professor at Western University.

A Collaborative Effort

The study harnessed the expertise of pathologists worldwide, who collectively scored an extensive image database containing over 1000 mitotic figures. High-quality images of mitotic figures were sourced from The Cancer Genome Atlas (TCGA). Each digitized image was annotated to ensure clarity and accuracy before being divided into tiles at 40x magnification for detailed examination.

“An international group of pathologists was recruited via social media to annotate and score these figures. The pathologists evaluated the different tiles to determine their consensus on whether each tile represented a mitotic figure,”

Dr. Cecchini explained.

He added that the most novel aspect of their work was the use of social media and publicly available datasets to rapidly recruit and coordinate a large international group of pathologists for the study.

“This innovative approach allowed for the swift generation of a comprehensive and diverse dataset of mitotic figures, which can be utilized to train AI systems in pathology,”

he said.

The agreement rates among pathologists were calculated to measure consistency. Statistical methods were used to assess inter-rater reliability, focusing on identifying patterns and discrepancies in scoring.

Findings

The 1000 Mitoses Project yielded several critical insights into the current state of mitotic figure classification. Pathologists demonstrated a median agreement rate of 80.2%, with individual scores ranging from 42.0% to 95.7%. The median agreement rate for individual mitotic figure tiles was higher, at 87.1%. This indicates that certain mitotic figures are more universally recognizable than others. The kappa statistic, a measure of agreement adjusted for chance, was calculated at 0.284, indicating fair agreement across all tiles.

“The findings emphasize the subjective nature of mitotic figure identification and the challenges in achieving consistent interpretations across different observers,”

noted Dr. Cecchini.

Notably, the study showed statistically significant differences in agreement rates when different cell cycle phases were considered. Mitotic figures in the prometaphase stage had the lowest agreement rates (median 81.2%), followed by mitotic figures in metaphase (median 89.4) and those in anaphase/telophase (median 85.9%). These findings suggest that mitotic figures in the prometaphase are more challenging to identify.

Moreover, the researchers found that nearly half (46.2%) of the tiles were labeled as potential mitotic figure mimics, with apoptotic bodies, karyorrhectic debris, and immune cells commonly misinterpreted as mitoses. These findings underscore the importance of developing robust strategies for differentiating true mitotic figures from mitotic figure mimics.

Implications

In addition to providing insights into the current state of mitotic figure classification, the dataset created in this study serves as an invaluable resource for training new pathologists, helping to standardize the identification process and reduce subjectivity. By highlighting the most challenging aspects of mitotic figure identification, this study provides a roadmap for the development of targeted educational tools to address these areas.

Furthermore, the dataset could serve as a valuable reference for the development of artificial intelligence (AI) algorithms that can assist pathologists in identifying mitotic figures with greater accuracy. AI tools trained on this comprehensive dataset could significantly enhance diagnostic accuracy and efficiency.

Looking Ahead

Although the 1000 Mitoses Project marks a significant advance toward the standardization of mitotic figure assessment, significant variability remains among pathologists, underscoring the inherent subjectivity in identifying mitotic figures. Moreover, the study highlights the difficulty in distinguishing mitotic figures from similar structures, such as apoptotic figures and tissue artifacts, which can confound results.

“The study highlighted several unanswered questions regarding the inherent variability in pathologist interpretations and the challenges of using whole slide images for mitotic figure identification. Future studies could focus on refining AI tools to better differentiate mitotic figures from their mimics in varied clinical settings and exploring novel education tools to train pathologists to better align on consensus mitotic figures,”

said Dr. Cecchini.

In addition, conducting longitudinal studies to assess the impact of improved mitotic figure classification on patient outcomes would provide valuable insight into the clinical benefits of these advancements. Moreover, continued development and refinement of AI tools using the dataset will be crucial. These tools should aim not only to assist in mitotic figure identification but also to integrate with other diagnostic parameters to provide comprehensive support for pathologists.

“The variability and subjectivity in identifying mitotic figures can significantly impact tumor diagnostics and clinical management. This underscores the need for more standardized and reliable methods, potentially through AI and machine learning tools, to assist pathologists in accurately classifying and counting mitotic figures, thereby improving diagnostic accuracy and patient outcomes,”

Dr. Cecchini concluded.

The study was supported by the Megan J. Davey Opportunity Fund and the Cancer Pathology Translational Research Grant from the Ontario Molecular Pathology Research Network and Ontario Institute for Cancer Research.

References

1. Lin S, Tran C, Bandari E, et al. The 1000 Mitoses Project: A Consensus-Based International Collaborative Study on Mitotic Figures Classification. Int J Surg Pathol. Published online April 16, 2024. doi:10.1177/10668969241234321