Towards trustworthy AI-driven leukemia diagnosis: A hybrid Hierarchical Federated Learning and explainable AI framework

Accurate detection and classification of microscopic cells from acute lymphoblastic leukemia remain challenging due to the difficulty of differentiating between cancerous and healthy cells. This paper proposes a novel approach to identify and categorize acute lymphoblastic leukemia that uses explain...

Full description

Saved in:
Bibliographic Details
Main Authors: Khadija Pervez, Syed Irfan Sohail, Faiza Parwez, Muhammad Abdullah Zia
Format: Article
Language:English
Published: Elsevier 2025-01-01
Series:Informatics in Medicine Unlocked
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S2352914825000061
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Accurate detection and classification of microscopic cells from acute lymphoblastic leukemia remain challenging due to the difficulty of differentiating between cancerous and healthy cells. This paper proposes a novel approach to identify and categorize acute lymphoblastic leukemia that uses explainable artificial intelligence and federated learning to train models across multiple institutions while keeping patient information decentralized and encrypted. The framework trains EfficientNetB3 for the classification of leukemia cells and incorporates explainability techniques to make decisions of the underlying model transparent and interpretable. The framework employs a hierarchical federated learning approach that allows distributed learning across clinical centers, ensuring that sensitive patient data remain localized. Explainability techniques such as saliency maps, occlusion sensitivity, and randomized input sampling for explanation with relevant evaluation scores are integrated in the framework to provide visual and textual explanations of model’s predictions to enhance interpretability. The experiments were carried out on a publicly available dataset consisting of 15,135 microscopic images. The performance of the proposed model was benchmarked against traditional centralized models and classical federated learning techniques. The proposed model demonstrated a 2.5% improvement in accuracy (96.5%) and a 5.4% increase in F1-score (94.4%) compared to baseline models. Hierarchical federated learning reduced communication costs by 15% while maintaining data privacy. The integration of explainable artificial intelligence improved the transparency of model decisions, with a high area under the ROC curve (AUC) of 0.98 for the classification of leukemia cells. These results suggest that the proposed framework offers a robust solution for intelligent systems for medical diagnostics and can also be extended to other medical imaging tasks.
ISSN:2352-9148