A new study aimed at automating breast cancer diagnostics included the development of an artificial intelligence (AI)-based system to predict patient survival outcomes.
The study by researchers at the University Medical Centre Utrecht demonstrated that an automated mitosis detection pipeline can match the prognostic accuracy of traditional light-microscope assessments of mitotic count (MC), a key marker in determining cancer aggression and patient prognosis.
Currently, pathologists assess MC by examining glass slides under a microscope, a process that relies on subjective interpretations. This can introduce variability between observers, impacting the accuracy and reproducibility of results, and complicating treatment decisions. The researchers noted that leveraging AI for MC assessment could offer a more consistent, reproducible, and efficient approach.
In the study which was published in The Journal of Pathology Clinical Research, the researchers trained a deep learning model on whole slide images of breast cancer tumors, developing it further to automatically detect and quantify mitotic hotspots—the areas of highest cell division. The AI model was based on technology from the 2021 Mitosis Domain Generalization Challenge (MIDOG21), designed to detect mitotic figures across various imaging conditions. This model was enhanced with an “automatic area selector” to identify the optimal area for mitotic counting within a standardized 2 mm² region, in line with international pathology guidelines.
The study included 912 breast cancer patients with long-term follow-up data from UMC Utrecht. The AI-based MC results were then compared to MC assessments previously documented in pathology reports. The findings show a high correlation between AI-based and traditional MC counts. This indicates that the AI system could offer a reliable alternative to human counts, the researchers wrote.
Both AI-derived and traditional MC assessments demonstrated similar prognostic power in predicting overall survival rates, both in univariate and multivariate analyses, with AI-based methods showing slightly better consistency. Additionally, the AI system is fully automated and cab be integrated into existing digital pathology systems. The tool has been adopted within the UMC Utrecht’s clinical workflow.
When pathologists reviewed and refined the AI-generated results in a subset of cases, the agreement with traditional methods further improved, showing the potential for AI to serve as a valuable support tool rather than a replacement. The research team made the AI model publicly available online for further development and validation across diverse datasets.
The study acknowledged the existing challenges in fully integrating AI-based MC into clinical settings, including the need for specialized IT infrastructure and ongoing training for healthcare staff. Further research is needed to examine the economic impact of AI in pathology labs in terms of time savings and productivity gains, the researchers noted.
The researchers declared no conflicts of interest.