To improve the classification of bone marrow cells by integrating federated learning and meta-learning to address data scarcity, privacy concerns, and enhance model adaptability and generalization capabilities.
Approach:
Methodology: Utilized a federated learning framework combined with a meta-learning algorithm based on prototypical networks, leveraging a ResNet-18 backbone.
Implementation: Conducted federated training using FedAvg across four clients, each with their own data.
Key Findings:
Achieved an accuracy of 96 ± 1% in classes with low sample sizes.
Demonstrated potential for high accuracy in unseen classes despite limited data.
Interpretation:
The proposed approach allows for effective model training while maintaining data privacy and addressing challenges related to data scarcity.
Limitations:
The study does not address the potential variability in model performance across different client datasets.
Challenges related to class imbalances and cytological heterogeneity remain.
Conclusion:
The federated learning and meta-learning approach shows promise for enhancing the classification of bone marrow smears while preserving data privacy.
