.png)
HER2 gene is amplified in up to 20% of invasive breast carcinomas. This amplification is associated with a poor prognosis as well as a best response to anthracycline-based chemotherapies and to HER2-targeting agents. HER2 amplification is closely linked to HER2 protein overexpression. Thus, as HER2-targeted therapy is exclusively effective in HER2-overexpressed and/or HER2-amplified breast carcinomas, precise assessment of HER2 status in routine practice is a critical step for the therapeutic decision making. We developed an artificial intelligence (AI) algorithm for HER2 protein overexpression quantification to aid pathologists with an accurate and reproducible assessment of HER2 immunohistochemistry scoring.
We developed and trained an end-to-end algorithm for scoring HER2 in breast carcinoma-stained whole slides images (WSI). The proposed algorithm automatically detects invasive regions of the tumor and allows the quantification and characterization of tumor cells on these zones. To achieve this, we propose a two-step framework: detecting invasive areas and identifying cells and membranes. First, 3878 square images (tiles) of size 512 pixels of HER2 IHC slides 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 identify the cells of interest. The assessment of the membrane completeness and intensity is also provided. Hence, this exhaustive quantification is used to assess the HER2 scores at the slide-level, in accordance with ASCO/CAP guidelines. The proposed pipeline provides pathologists with a readily interpretable score along with the details of the slide analysis. In addition, our algorithm works in an end-to-end manner, without requiring any assistance from pathologists. The performance of the model is investigated over a left-out dataset and compared to the ground-truth.
The model’s performance in predicting HER2 scores was evaluated on a left-out dataset of 56 HER2 IHC slides previously scored by pathologists. The test set comprised slides distributed across the four HER2 categories: 0 (21%), 1+ (36%), 2+ (29%), and 3+ (14%). The model demonstrated an overall accuracy of 82.1% in correctly classifying HER2 scores. Notably, the model excelled in accurately identifying the two most distinct scores (0 and 3+), achieving a 100% recall rate for both categories The pipeline is also challenged using external datasets of breast cancer, which were independently screened by experienced pathologists.
Our end-to-end approach addresses the lack of interpretability in AI-based WSI processing and eliminates the need for pathologists to perform additional tasks, such as selecting regions of interest. Our solution automatically performs exhaustive tumor cell analysis in invasive tissue regions and improves HER2 Score interpretation with detailed quantification. The promising results achieved at this stage of the development pave the way for further enhancements that would significantly improve the accuracy of HER2 Scoring performed by AI along with the inter-observer agreement among pathologists assisted by our algorithm.
Florian Thomas, Yahia Salhi, Celine Bossard, Benedicte Cormier, Bruno Poulet, Laetitia Lambros, Ilham Chokri, Frederique Jossic, Soumanth Thanguturi, Sanae Salhi, Jerome Chetritt, Isabelle de Pinieux