Hierarchical Swin Transformer Ensemble with Explainable AI for Robust and Decentralized Breast Cancer Diagnosis

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Computational Pathology

Background:

• Early breast cancer detection significantly reduces mortality.
• Deep learning models face challenges in healthcare, including data privacy, overfitting, and interpretability.
• Federated learning (FL) offers a privacy-preserving approach for collaborative AI model training.

Purpose of the Study:

• To develop a robust and privacy-preserving AI system for breast cancer diagnostics.
• To address limitations of traditional deep learning models in healthcare settings.
• To enhance the accuracy and interpretability of AI-driven breast cancer detection.

Main Methods:

• Proposed BreastSwinFedNetX, a federated learning (FL)-enabled ensemble system.
• Combined four hierarchical Swin Transformer variants (Tiny, Small, Base, Large) with a Random Forest (RF) meta-learner.
• Utilized FL for decentralized training, preserving data locality and privacy.
• Integrated Explainable AI (XAI) using Grad-CAM for feature visualization.

Main Results:

• Achieved high performance across five benchmark datasets: BreakHis, BUSI, INbreast, CBIS-DDSM, and a Combined dataset.
• Reported an F1 score of 99.34% on BreakHis, PR AUC of 98.89% on INbreast, and MCC of 99.61% on the Combined dataset.
• Demonstrated strong generalization and robustness through extensive ablation studies.

Conclusions:

• BreastSwinFedNetX offers a scalable and trustworthy AI solution for real-world breast cancer diagnostics.
• The FL approach ensures data privacy and security, compliant with GDPR and HIPAA.
• The integration of XAI enhances transparency and facilitates clinical adoption.