Deep-Learning-Based Approach in Cancer-Region Assessment from HER2-SISH Breast Histopathology Whole Slide Images.

Area of Science:

• Digital Pathology
• Computational Biology
• Oncology

Background:

• Fluorescence in situ hybridization (FISH) is the gold standard for HER2 testing in breast cancer but has limitations.
• Silver-enhanced in situ hybridization (SISH) offers an automated, bright-field microscopy-compatible alternative.
• Accurate HER2 status determination is crucial for guiding breast cancer treatment decisions.

Purpose of the Study:

• To develop and validate a deep learning model for classifying HER2-SISH whole slide images (WSIs).
• To automate the identification and localization of Normal, Amplified, and Non-Amplified regions in HER2-SISH WSIs.
• To provide an efficient computational tool for pathologists in assessing HER2 status.

Main Methods:

• A two-stage deep learning approach was employed, starting with region-level model evaluation.
• A Vision Transformer (ViT) model, utilizing transfer learning, was applied to HER2-SISH WSIs.
• Pseudo-color maps were generated and overlaid onto WSIs for visualization of classified regions.

Main Results:

• The model achieved 99.9% patch-level classification accuracy and 78.8% generalization accuracy on a private dataset.
• K-fold cross-validation demonstrated robust performance with an average accuracy of 98%.
• The method successfully identified and localized HER2-amplified regions within complex WSIs.

Conclusions:

• Deep learning, specifically using a ViT model, can accurately classify HER2 status in SISH images.
• This automated approach shows significant promise for clinical pathology, aiding in efficient HER2 assessment.
• The developed method has the potential to enhance diagnostic workflows and improve breast cancer treatment outcomes.