Smart FL: meta-learning for federated blood marrow smear classification

Accurate classification of bone marrow cells is crucial for diagnosing hematological disorders, yet integrating machine learning and deep learning models into clinical practice faces hurdles like data privacy and data scarcity. Existing models often lack the adaptability that is important in the medical field, limiting their practicality. This study introduces an approach using federated learning and meta-learning to tackle these challenges. The objectives for this study were multifaceted. Firstly, it address the adverse effects of data scarcity commonly found in medical datasets due to which models are unable to predict classes which do not have enough examples. Secondly, this study tackles the privacy concerns by adopting a federated learning framework, allowing model training without centralizing sensitive data. Additionally, the goal is to make the federated process personalized to each client for enhancing individual accuracy. Lastly, this study seeks to improve the generalization capabilities of the models, enabling robust performance across diverse patient populations. The proposed approach involves leveraging a ResNet-18 backbone for the meta-learning algorithm, specifically one based on prototypical networks. This framework is then implemented and tested in a federated manner using FedAvg across four clients, each possessing their own data. The main focus of this research is to achieve high accuracy, particularly in unseen classes with limited samples. This approach yields promising results, achieving an accuracy of 96 ± 1% in classes with low sample sizes. Through this approach, hospitals, clinics, and researchers can make sure that their data remains private while being able to benefit from deep learning research.