Evaluation of the mitotic score of invasive breast carcinomas on digital slide: development and contribution of a mitosis detection algorithm

Introduction: Nottingham grading system is a major prognostic factor for invasive breast carcinoma (IBC). Its determination requires the evaluation of the mitotic score (MS) which is subject to low intra- and inter-observer reproducibility. The MS shall be performed in the most proliferative area of the tumor, which determination is hard but critical. Artificial intelligence based tools could help pathologists to detect mitosis on whole slide images (WSI). Objective: The aim of this study was to evaluate the contribution of a mitosis detection algorithms pecifically developed to assist the pathologist during the evaluation of the MS on WSI. Methods: Algorithmic mitosis detection is a two-step process: first the algorithm detects candidate objects resembling mitosis, then the selection is refined by a classifier. The densest mitoticregions are shown to the pathologist, then he can establish the MS with algorithm results. For this study, three expert pathologists have determined a consensual ground truth for MS on fifty WSI of IBC. Those slides were also submitted to two readers pathologists who evaluated the MS of each slide twice, with and without the assistance of the algorithm, with a four week wash-out period. Interobserver reproducibility was measured by evaluating the scores obtained with, and without assistance between two readers pathologists and was also measured between each reader pathologist and the expert ground truth to determine the accuracy of the established score. Results:Baseline linearly weighted Cohen's Kappa for interobserver agreement of MS between two readers pathologists was 0.482. Using the algorithm generated mitotic detection in WSI, the agreement score increased to 0.672. Baseline linearly weighted Cohen's Kappa for interobserver agreement of MS between each reader pathologist and expert consensus was 0.378 and 0.457 for pathologist 1 and 2 respectively. Using the algorithm generated mitoticdetection in WSI, the agreement score increased respectively to 0.629 and 0.726. Conclusion:The use of the developed algorithm constitutes a viable approach to assist the pathologist for the evaluation of the MS of IBC on WSI. Its use makes it possible to improve interobserver reproducibility between pathologists and the accuracy of the score established by expert consensus. The use of such a tool constitutes a new approach in the evaluation of the mitoticscore which could lead to an evolution of practices.