Conventional Machine Learning versus Deep Learning for Magnification Dependent Histopathological Breast Cancer Image...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Breast cancer diagnosis relies heavily on histopathological image analysis.
• Existing computer-aided diagnosis (CAD) methods often lack interpretability.
• There is a need for explainable AI in medical diagnostics.

Purpose of the Study:

• To compare the performance of conventional machine learning (CML) and deep learning (DL) methods for breast cancer classification.
• To provide visual interpretations of classification decisions for improved clinical trust.
• To evaluate the efficacy of DL models using attention maps for explainability.

Main Methods:

• Handcrafted feature extraction and classical classifiers for CML methods.
• Transfer learning with the VGG-19 deep learning architecture for DL methods.
• Evaluation on the BreaKHis and KIMIA Path960 datasets, with attention map visualization for DL interpretability.

Main Results:

• Deep learning methods outperformed conventional machine learning approaches.
• DL achieved accuracies between 85.65%-89.32% (binary) and 63.55%-69.69% (eight-class).
• Visual explanations using attention maps enhanced the interpretability of DL models.

Conclusions:

• Deep learning models offer superior performance for breast cancer classification from histopathological images.
• Visual interpretability through attention maps increases pathologist trust in AI diagnostic tools.
• Explainable AI holds significant promise for supporting clinical decision-making in breast cancer diagnosis.