Weakly-supervised deep learning models enable HER2-low prediction from H &E stained slides.

Area of Science:

• Oncology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• HER2-low breast cancer is a newly identified subtype with therapeutic implications.
• Current HER2 assessment involves multiple tests, including immunohistochemistry and in situ hybridization, which can be time-consuming and costly.
• There is a need for efficient and cost-effective methods for HER2 status determination.

Purpose of the Study:

• To develop and evaluate a deep learning model for predicting HER2-low breast cancer status directly from histopathological images.
• To explore the performance of different classification models in distinguishing HER2-negative, HER2-low, and HER2-high breast cancer subtypes.
• To utilize explainable AI to identify histological patterns associated with HER2 subtypes.

Main Methods:

• A self-supervised, attention-based, weakly supervised learning method was employed.
• Six distinct classification models were trained using 1437 histopathological images from 1351 breast cancer patients.
• An attention-based model was used to visualize and understand the regions of interest in the decision-making process.

Main Results:

• The study highlights the critical dependence of model performance on the reliability of assay-based HER2 testing for training data.
• Explainable AI techniques successfully identified specific histological patterns linked to different HER2 subtypes.
• The developed models demonstrated the potential for accurate HER2 subgroup classification.

Conclusions:

• Deep learning technologies can be effectively applied to classify HER2 subgroup statuses in breast cancer.
• This AI-driven approach has the potential to enhance the clinical decision-making toolkit for oncologists.
• The findings pave the way for faster and potentially more cost-effective HER2 assessment in breast cancer diagnostics.