Automated Quantification of HER2 Amplification Levels Using Deep Learning.

Area of Science:

• Oncology
• Biotechnology
• Medical Imaging

Background:

• Accurate HER2 (Human Epidermal growth factor Receptor 2) amplification assessment is crucial for selecting patients eligible for anti-HER2 targeted therapies in breast and gastric cancers.
• Manual evaluation of HER2 amplification using fluorescence in situ hybridization (FISH) and dual in situ hybridization (DISH) is time-consuming, labor-intensive, and prone to subjective errors due to image complexities like blurry boundaries and overlapping cells.

Purpose of the Study:

• To develop and validate a novel deep learning model for automated and objective quantification of HER2 amplification from FISH and DISH images.
• To improve the accuracy and efficiency of HER2 assessment for precise patient stratification in anti-HER2 therapy.

Main Methods:

• A soft-sampling cascade deep learning model and a signal detection model were developed to quantify CEN17 and HER2 signals in individual cells.
• The model performs instance segmentation of HER2-amplified cells, requiring co-localization of both CEN17 and HER2 signals.

Main Results:

• The proposed deep learning model achieved high accuracy, recall, and F1-score in instance segmentation of HER2-related cells on both FISH and DISH datasets.
• The model significantly outperformed seven state-of-the-art deep learning methods.
• When applied to gastric cancer HER2 DISH assessment, the model demonstrated promising predictive performance with 97.67% accuracy and 96.15% precision.

Conclusions:

• The developed deep learning approach offers an accurate, objective, and efficient method for HER2 amplification assessment, overcoming limitations of manual evaluation.
• This automated tool can significantly aid in patient selection for HER2-targeted therapies in breast and gastric cancers, potentially improving treatment outcomes.