FL-QNNs: Memory Efficient and Privacy Preserving Framework for Peripheral Blood Cell Classification

Human blood predominantly consists of plasma, erythrocytes, leukocytes, and thrombocytes. It is crucial for the transportation of oxygen and nutrients and for storing the health information of the human body. The body uses blood cells to fight against diseases and infections. Consequently, blood analysis enables clinicians to evaluate a person’s physiological state. These blood cells are categorized into eight classes such as neutrophils, eosinophils, basophils, lymphocytes, monocytes, immature granulocytes, erythroblasts, and platelets based on nuclear characteristics, morphology, and cytoplasmic composition. Historically, pathologists and hematologists have discovered and analyzed specimens using microscopy before physical categorization, a method that is time consuming and susceptible to human error. Consequently, automating this procedure is imperative. This study proposes a resource efficient, privacy preserving, optimized memory framework by incorporating two approaches: Federated learning and quantized neural network (FL-QNNs) for peripheral blood cell (PBC) image classification. The data set include the total 17,092 images of eight classes. The study began by training the model with seven Deep learning Convolutional neural network (DL-CNN) model such as LeNet, SimpleCNN, MiniAlexNet, SmallVGG, TinyResNet, MiniMobileNet, and Multiple layer perceptron (MLP) and consider them as a baseline models achieving a 98-99% accuracy. Then implement the federated learning (FL) framework with IID (Independent and identically distributed) and Non-IID datasets maintain similar results by preserving data. Further FL-QNNs approach is used with these baseline models and achieved 98-99% accuracy and while optimizing memory, preserving data and resources. This makes the technique exceptionally appropriate for implementation in real-world, privacy-sensitive, and resource-limited healthcare settings. FL-QNN is recognized as an effective approach for minimizing substantial computational resources and memory associated with several Deep Learning (DL) methods.