DEEP LEARNING ARTIFICIAL INTELLIGENCE-BASED HER 2 IMMUNOHISTOCHEMISTRY SCORING FOR BREAST CANCER USING WHOLE SLIDE IMAGES: A DECISION SUPPORT SYSTEM (HER2-AID STUDY) ganeshbhat92@gmail.com

 

DEEP LEARNING ARTIFICIAL INTELLIGENCE-BASED HER 2 IMMUNOHISTOCHEMISTRY SCORING FOR BREAST CANCER USING WHOLE SLIDE IMAGES: A DECISION SUPPORT SYSTEM (HER2-AID STUDY)

2

 

 

 

 

India

Oral or Poster

Immunohistochemistry (IHC) is essential for breast cancer diagnosis and management. Manual IHC scoring by pathologists is time-consuming, prone to subjectivity, and may lead to inter-observer variability. This study presents Her2-AID (HER2-Artificial Intelligence-based Immunohistochemical Decision support system), a deep learning tool for automated HER2 IHC scoring using Whole Slide Images (WSIs).

A convolutional neural network based on the pretrained DenseNet-121 architecture was developed using the publicly available Breast Cancer Immunohistochemical (BCI) dataset containing ~5000 images representing various HER2 expression levels. WSIs were tiled into 256×256 pixel patches, and the dataset was split 80:20 for training and validation while maintaining class balance. The DenseNet model, pretrained on ImageNet, was fine-tuned for this task. Patch-level predictions were aggregated using majority voting to obtain slide-level scores via majority voting and compared to grouth truth labels. Performance was evaluated using confusion matrices.

The Her2-AID model achieved 63% accuracy at the patch level, which improved to 78% with aggregated slide-level predictions. It performed well in distinguishing strong (3+) and negative (0) HER2 expressions, with moderate accuracy in equivocal (1+/2+) categories. Enhancements through pathologist-curated image selection, data augmentation, and fine-tuning of DenseNet layers improved model performance.

This study highlights the feasibility and effectiveness of using DenseNet-based deep learning for HER2 IHC scoring in breast cancer. With further training on larger datasets, the model could become a scalable, reproducible tool to aid pathologists in making faster and more consistent diagnostic decisions.