A personalized communication efficient federated learning framework with low rank adaptation for intelligent leukemia diagnosis

Leukemia diagnosis with medical imaging necessitates the development of highly accurate and individualized models that uphold data privacy among institutions. This research proposes a framework named FedPerLoRA-Health, a communication-efficient federated learning framework that combines federated personalization and low rank adaptation with EfficientNet architectures for personalized leukemia detection. The proposed PerFLR-EffNet algorithm holds the structural efficiency of EfficientNet variants B0 and B2 as backbone models, facilitating parameter-efficient updates and local personalization across diverse client datasets. Within this framework, personalized layers undergo local training, whereas LoRA-adapted global layers are disseminated to reduce communication overhead. The proposed method is assessed on a Blood Cells Cancer Acute Lymphoblastic Leukemia (ALL) dataset with classification-based metrics such as accuracy, precision, recall and F1-score and federated learning-based metrics such as communication cost and convergence rate. The efficiency of the proposed model is analysed by comparing it with the baseline models such as centralized EfficientNetB0 and EfficientNetB2 without personalized Federation. Experimental results indicate that PerFLR-EffNet attains a better average classification accuracy of 98.67% and also proves to be communication efficient by reporting reduced number of trainable parameters and a reduction in communication overhead by 88.12% when compared with the baseline models.