Automated Ki-67 labeling index assessment in prostate cancer using artificial intelligence and multiplex fluorescence immunohistochemistry

The Ki-67 labeling index (Ki-67 LI) is a strong prognostic marker in prostate cancer, although its analysis requires cumbersome manual quantification of Ki-67 immunostaining in 200-500 tumor cells. To enable automated Ki-67 LI assessment in routine clinical practice, a framework for automated Ki-67 LI quantification, which comprises three different artificial intelligence analysis steps and an algorithm for cell-distance analysis of multiplex fluorescence immunohistochemistry staining (mfIHC), was developed and validated in a cohort of 12,475 prostate cancers. The prognostic impact of the Ki-67 LI was tested on a tissue microarray (TMA) containing one sample each patient. A "heterogeneity TMA" containing 3 to 6 samples from different tumor areas each patient was used to model Ki-67 analysis of multiple different biopsies and 30 prostate biopsies were analyzed to compare a "classical" bright field-based Ki-67 analysis with the mfIHC-based framework. The Ki-67 LI provided strong and independent prognostic information in 11,845 analyzed prostate cancers (p<0.001 each) and excellent agreement was found between the framework for automated Ki-67 LI assessment and the manual quantification in prostate biopsies from routine clinical practice (intraclass correlation coefficient: 0.94 [95% CI: 0.87 - 0.97]). The analysis of the heterogeneity TMA revealed that the Ki-67 LI of the sample with the highest Gleason score (AUC:0.68) was as prognostic as the mean Ki-67 LI of all six foci (AUC:0.71 [p=0.24]). The combined analysis of the Ki-67 LI and Gleason score obtained on identical tissue spots showed that the Ki-67 LI added significant additional prognostic information in case of classical ISUP grades (AUC:0.82 [p=0.002]) and quantitative Gleason score (AUC:0.83 [p=0.018]). The Ki-67 LI is a powerful prognostic parameter in prostate cancer, which is now applicable in routine clinical practice. In case of multiple cancer positive biopsies, the sole automated analysis of the worst biopsy was sufficient.