Enhancing histopathological image classification of invasive ductal carcinoma using hybrid harmonization techniques

Area of Science:

• Pathology
• Medical Imaging
• Computational Biology

Background:

• Histopathological image analysis for breast cancer diagnosis faces challenges due to intra- and inter-center variability.
• Variability arises from image acquisition differences and diverse patient clinical characteristics, impacting model robustness.

Purpose of the Study:

• To develop a robust feature engineering-based pipeline for classifying invasive ductal carcinomas and benign tumors in histopathological images.
• To address and rectify batch effects and data variability within and between different data centers.
• To enhance the accuracy and efficiency of breast cancer diagnostic models.

Main Methods:

• A feature engineering pipeline involving feature extraction.
• Implementation of multiple harmonization techniques to correct for intra- and inter-center batch effects.
• Performance evaluation on public breast cancer datasets (BreaKHIS and IDCDB) using accuracy, precision, and recall metrics.

Main Results:

• The proposed pipeline achieved 90-95% accuracy in classifying benign and malignant tumors.
• Harmonization significantly improved performance for classifying data from different databases.
• Top model performance reached 94.7% on IDCDB and 95.2% on BreaKHis, outperforming existing methods.

Conclusions:

• The feature-engineering-based pipeline effectively classifies tumors while mitigating variability issues.
• Harmonization techniques are crucial for improving the learning and testing performance of classification models.
• The pipeline shows potential for enhancing breast cancer diagnosis and can be adapted for other diseases.