Citation

Nabi, Md Serajun (2025) Automated scoring of HERS2 immunohistochemistry breast tumor biomarker. Masters thesis, Multimedia University.

Full text not available from this repository.

Abstract

Breast cancer remains one of the top causes of death among women globally. Proper assessment of the human epidermal growth factor receptor 2 (HER2) biomarker is crucial for deciding who can receive targeted anti-HER2 therapies. HER2 immunohistochemistry (IHC) is the main screening method used in pathology labs. However, its manual interpretation can be subjective and often varies between different observers, especially in borderline categories (1+ and 2+), where treatment options can be unclear. These issues highlight the need for more automated, consistent, and understandable HER2 scoring methods. This thesis presents a deep learning and explainable AI (XAI) framework for automated HER2 IHC scoring using high-resolution whole slide images (WSIs). The process begins with creating a dataset from 107 HER2-stained WSIs collected from the University of Malaya Medical Centre. This is followed by annotating regions of interest (ROI), extracting patches, and filtering based on HSV quality. Two patch-based models were developed and assessed: a custom Convolutional Neural Network (CNN) and a fine-tuned DenseNet121. These models laid a solid foundation for supervised learning. To improve robustness, a ResNet50 model was introduced with a hybrid layer-unfreezing strategy. It was trained on three datasets (HER2-IHC-40x-WSI, HER2-IHC-40x-Patch, and BCI-10×), allowing for assessments of performance across different magnifications and datasets. Post-hoc interpretability measures were added using SHAP and Grad-CAM for the CNN and DenseNet121 models, along with ScoreCAM for ResNet50 to create membrane-level attention maps that align with pathologist expectations. Quantitative experiments show that both the CNN and DenseNet121 models achieve 93% accuracy. DenseNet121 has smoother convergence and better spatial localization, while the CNN offers better pixel-level interpretability and lower training loss. The ResNet50 model achieved the best overall performance with 96% accuracy. It also showed a slight performance drop between magnifications (∆F1 < 0.04 between 10× and 40×) and performed well in external validation on the BCI dataset. The evaluation of explainability indicated high membrane activation precision (MAP = 79%) and consistent alignment with areas annotated by pathologists, outperforming previous Grad-CAM-based studies that reported much lower spatial alignment. In conclusion, this multi-model, multi-resolution framework provides high accuracy, better robustness, and meaningful interpretability for HER2 IHC scoring. These results advance the development of reliable AI-assisted diagnostic tools in digital pathology and demonstrate the potential for use in real-world clinical settings to ensure consistent and clear HER2 evaluation.

Item Type:	Thesis (Masters)
Additional Information:	Call No.: Q325.73 .M37 2025
Uncontrolled Keywords:	Deep learning (Machine learning)
Subjects:	Q Science > Q Science (General) > Q300-390 Cybernetics
Divisions:	Faculty of Artificial Intelligence & Engineering (FAIE)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	19 Jan 2026 03:50
Last Modified:	19 Jan 2026 03:50
URII:	http://shdl.mmu.edu.my/id/eprint/15188

Downloads

Edit (login required)