2329P - Evidence for the utility of artificial intelligence and image analysis in Ki-67 quantification in solid tumors

Background

Although it is an important biomarker in oncology, Ki-67 immunohistochemistry analysis has yet to be standardized. Working groups have provided guidelines for Ki-67 scoring in different cancer types to limit pathologist’s variability. Digital analysis solutions to assist scoring with image analysis or AI have recently emerged in the evaluation of Ki-67 as rapid and robust solutions. In this context, we compared the results of Ki-67 scoring performed with Aiforia® (AI platform) and Halo® (image analysis supervised software) against two independent pathologists on solid tumors.

Methods

We stained 192 tumors of various origins including breast and prostate with Ki-67 (clone 30-9). Pathologists were trained following the International Ki-67 Working Group recommendations and scored tissues accordingly. Based on deep learning, Aiforia was able to automatically score Ki-67 positive tumor cells within minutes. The random forest classifier from Halo software was used to separate the image into tumor, non-tumor and background. After cell segmentation, Ki-67 positivity was assessed by thresholding.

Results

Our study shows a very high consistency of results obtained for Ki-67 scoring between the image analysis softwares, Aiforia and Halo (r²=0.94), on all solid tumors. The correlation obtained between the two pathologists was weaker (r²=0.8), despite the training and guidelines, but remains within an acceptable range. Of the 18 types of primary tumors analyzed, only 2 (thyroid and stomach) showed correlations of less than 0.75 between the image analysis softwares with inter-pathologists' correlation still high (0.73 and 0.86). Depending on the organ, there were significant variations between pathologists, whereas the scores obtained with the image analysis softwares remain very close. For example, with prostate (n=8), a 0.81 correlation for Ki-67 scores between Aiforia and Halo was found while the inter-pathologists' correlation was 0.68.

Conclusions

This work shows that AI based image analysis tools such as Aiforia and image analysis supervised software provide valuable assistance in IHC based clinical analysis, drastically reducing inter-pathologist variability or as an arbitration tool in Ki-67 scoring of solid tumors.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.