Privacy-preserving multi-class classification of acute lymphoblastic leukemia subtypes using federated learning

Early and accurate diagnosis for a highly aggressive hematological malignancy: Acute Lymphoblastic Leukemia. This is where automated, privacy-preserving diagnostic solutions can not only ease the burden of current diagnostic approaches but also avoid invasive, time-intensive, and prone to error. In this study, we present a Federated Learning framework for the Multi-Class classification of Acute Lymphoblastic Leukemia subtypes based on Peripheral Blood Smear images. To deal with class imbalance, data augmentation techniques were applied, and then pre-trained convolution neural networks such as InceptionV3, DenseNet121, and Xception were fine-tuned to extract features. Of these, InceptionV3 performed the best with an accuracy of 95.49% in the Federated Learning framework guaranteeing the privacy of patient data through differential privacy mechanisms. Through comparative analysis, it was confirmed that in using the Federated Learning approach, the high diagnostic accuracy and robust generalization against different datasets were preserved, while outperforming centralized learning. By proposing a scalable, privacy-compliant solution for all diagnoses, Acute Lymphoblastic Leukemia diagnoses may be transformed into the new practice of hematological oncology.