.png)
Mitosis count is a critical factor in the grading of breast cancer, directly influencing treatment decisions and patient outcomes. However, manual detection and quantification of mitoses in histopathological images are time-consuming and subject to inter-observer variability. This study presents an AI-based system, trained on annotated histological images, is designed to assist pathologists in accurately detecting mitotic figures in invasive areas of breast cancer carcinomas samples and assess its analytical and clinical performances.
An analytical comparison of the mitoses detected by the AI model was performed by confronting them with those independently identified by two pathologists across 20 regions of interest on 20 WSI. In a second step, a mutual review of the mitoses detected by both pathologists was conducted to establish a consensus. The pathologists then compared the mitoses detected by the AI model with those from their consensus. This approach allows us, first, to measure the concordance between the pathologists, then between the pathologists and the AI model, and finally, to compare the consensus mitosis count (gold standard) with those detected by the AI. Additionally, a clinical evaluation was conducted on a cohort of 122 breast cancer patients with a five-year follow-up to assess the prognostic power of the AI-generated mitotic scores, using the hazard ratio (HR) from a Cox univariate regression model.
The pathologists identified 240 and 400 mitoses, respectively, with 208 in common. Upon reviewing independently the mitoses proposed by the AI, 88% were validated as correct, increasing the number of agreed-upon mitoses by 41%, resulting in a total of 295 mitoses. Each pathologist added 24% and 28% more mitoses to their original counts based on the AI's findings. After a consensus by pathologists, the AI model achieved a precision of 88% and a recall of 71%. Finally, when the AI-generated mitotic score was added to a Cox model, it demonstrated a significant prognostic impact on patient outcomes (HR 1.32, p=0.01).
This study highlights the potential of AI-assisted mitosis detection to improve the accuracy and consistency of breast cancer grading. The AI model showed strong concordance with pathologists and significantly increased mitotic detection rates. Additionally, the AI-derived mitotic score proved to be a valuable prognostic factor, with a strong association to patient outcomes, underscoring the clinical relevance of AI in diagnostics.
Baptiste Gourdin: None; Julia Ratour: None; Jean-François Jazeron: None; Céline Bossard: None; Yahia Salhi: None; Jérôme Chetritt: None