← ResourcesPublications
Training and Validating an Artificial Intelligence Algorithm for HER2 Assessment in Breast Carcinoma-Stained Whole-Slide Images
01 MAR 2024
Abstract
Background
HER2 amplification occurs in up to 20% of invasive breast carcinomas, impacting prognosis and therapeutic response. Thus, accurate assessment of HER2 status is vital for treatment decisions. We developed a fully-automated AI algorithm for precise HER2 protein overexpression quantification, aiding pathologists in immunohistochemistry scoring.
Methods
We developed and trained an end-to-end algorithm for scoring HER2 in breast carcinoma whole slides images (WSI) stained with immunohistochemistry. The proposed algorithm automatically detects invasive tumor regions and allows the quantification and characterization of tumor cells on these zones. To achieve this, 3878 square images of size 512 pixels of HER2-immunostained WSI were labeled at magnification x5 and used to train a deep learning model able to detect invasive regions. Hence, 4793 tiles were annotated at the cell level, at magnification x40, and used to train another model to segment tumor cells as well as to analyse the cell membrane. Hence, this exhaustive quantification was used to assess the HER2 score in the WSI, in accordance with ASCO/CAP guidelines. The proposed pipeline provides an interpretable score along with the details of the slide analysis without requiring any annotations from pathologists. The performance of the model was investigated over a left-out dataset comparing it to ground-truth scores. The ground truth scores were determined by majority voting after an independent scoring made by three senior pathologists. Besides, after a washout period of two months, the slides were rescored with the AI Assistance to compare the inter-observer agreement between these two steps.
Results
The model exhibited an overall balanced accuracy of 90.0% on the testing dataset, thus demonstrating its capability to assist pathologists in their routine practice while providing an interpretable HER2 Score. Notably, 100% of the HER2 3+ scores and 93.3% of the HER2-Low cases were identified by the model, yielding promising results to identify patients eligible to targeted therapy. Moreover, the inter-observer agreement rose from 57% to 75% with the assistance of AI (Fig. 3).
Conclusion
This study demonstrates the efficacy of the proposed fully-automated AI system in accurately assessing HER2-immunostained Whole Slide Images, in accordance with the 2018 ASCO/CAP Guidelines. By providing pathologists with easily interpretable scores, such an AI tool could be used as a regular aid in clinical decision-making, improving reproducibility and aiding in the assessment of challenging cases due to the detailed information it provides for each cases. Significantly, the introduced model exhibited reliability and consistency particularly in HER2-Low cases, setting a foundation for future advancements in precision medicine and diagnostic accuracy within HER2-related pathology.