Attention-Based Deep Learning Approach for Breast Cancer Histopathological Image Multi-Classification.

Area of Science:

• Digital pathology
• Computational biology
• Medical imaging analysis

Background:

• Breast cancer diagnosis from histopathology images is challenging due to time constraints and potential human error.
• Deep learning offers improved accuracy and efficiency but faces limitations with small datasets and subtle variations.
• Attention mechanisms can enhance feature representation in deep learning models.

Purpose of the Study:

• To propose the Efficient Channel Spatial Attention Network (ECSAnet) for accurate and efficient breast cancer classification.
• To evaluate the performance of ECSAnet against established deep learning architectures.
• To assess the impact of attention mechanisms and stain normalization on diagnostic accuracy.

Main Methods:

• Developed ECSAnet by integrating EfficientNetV2 with a Convolutional Block Attention Module (CBAM).
• Fine-tuned ECSAnet on the BreakHis dataset using Reinhard stain normalization and image augmentation.
• Compared ECSAnet's performance against AlexNet, DenseNet121, EfficientNetV2-S, InceptionNetV3, ResNet50, and VGG16.

Main Results:

• ECSAnet achieved high classification accuracies across different magnifications: 94.2% (40×), 92.96% (100×), 88.41% (200×), and 89.42% (400×).
• ECSAnet demonstrated superior performance compared to other benchmark models in most tested settings.
• The study confirmed the effectiveness of CBAM in boosting classification accuracy and the importance of stain normalization.

Conclusions:

• ECSAnet represents a significant advancement in deep learning for breast cancer histopathology image analysis.
• The integration of attention mechanisms and proper data preprocessing are crucial for robust diagnostic tools.
• The findings suggest ECSAnet's potential to improve the efficiency and reliability of breast cancer diagnosis.