Immunohistochemistry for nuclear Ki67 in breast cancer is ofinterest for patient management, but interlaboratory variabilityhas limited its clinical implementation and impact. Rimm et al.investigated whether automated assessment of Ki67 index utilizingmachine learning approaches can achieve superior reproducibility.They used seven unique scanning devices and 10 differentsoftware packages. While their primary endpoint was concordanceof results across all solutions, they also examined variability whensame solutions were used. Measures of average and maximalscores were compared, and the former were more reproducible.The correlation coefficient for all solutions was 0.83; this improvedto 0.89 for the solution deployed at multiple sites, which wassimilar to that obtained using the same slides in a pathologist-readstudy (0.87). These automated approaches are showing promise,and with some refinement and additional confirmation of clinicalvalidity and utility could be deployed in clinical practice.
展开▼
