Explainable federated transformer framework for joint leukemia classification and stage prediction.

Area of Science:

• Artificial Intelligence in Oncology
• Medical Image Analysis
• Natural Language Processing in Healthcare

Background:

• Leukemia diagnosis relies on analyzing hematological images and clinical reports.
• Existing machine learning models are often unimodal, centralized, and lack clinical interpretability.
• Decentralized data integration and privacy preservation are significant challenges in medical AI.

Purpose of the Study:

• To propose a federated multimodal architecture for joint leukemia diagnosis and staging.
• To integrate Vision Transformers (ViT) for image encoding and ClinicalBERT for text classification.
• To enable privacy-preserving, interpretable, and accurate leukemia diagnosis on decentralized medical devices.

Main Methods:

• Developed a federated multimodal architecture combining ViT and ClinicalBERT.
• Implemented a cross-modal fusion layer to synthesize image and text data into a shared semantic space.
• Utilized a federated learning protocol for decentralized data analysis and SHAP for explainability.

Main Results:

• The proposed system achieved higher accuracy and F1-score compared to unimodal and centralized baselines.
• Demonstrated interpretable and patient-specific explanations for both visual and textual data.
• Showcased robustness across non-IID data distributions and scalability in simulated healthcare networks.

Conclusions:

• The federated multimodal architecture offers a promising solution for accurate, private, and interpretable leukemia diagnosis.
• The system's design aligns with clinical expectations and is suitable for real-world deployment in diagnostic oncology.
• This approach addresses limitations of current AI models by enabling decentralized learning and enhancing transparency.